tools/perf/builtin-c2c.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * This is rewrite of original c2c tool introduced in here:
   4  *   http://lwn.net/Articles/588866/
   5  *
   6  * The original tool was changed to fit in current perf state.
   7  *
   8  * Original authors:
   9  *   Don Zickus <dzickus@redhat.com>
  10  *   Dick Fowles <fowles@inreach.com>
  11  *   Joe Mario <jmario@redhat.com>
  12  */
  13 #include <errno.h>
  14 #include <inttypes.h>
  15 #include <linux/compiler.h>
  16 #include <linux/err.h>
  17 #include <linux/kernel.h>
  18 #include <linux/stringify.h>
  19 #include <linux/zalloc.h>
  20 #include <asm/bug.h>
  21 #include <sys/param.h>
  22 #include "debug.h"
  23 #include "builtin.h"
  24 #include <perf/cpumap.h>
  25 #include <subcmd/pager.h>
  26 #include <subcmd/parse-options.h>
  27 #include "map_symbol.h"
  28 #include "mem-events.h"
  29 #include "session.h"
  30 #include "hist.h"
  31 #include "sort.h"
  32 #include "tool.h"
  33 #include "cacheline.h"
  34 #include "data.h"
  35 #include "event.h"
  36 #include "evlist.h"
  37 #include "evsel.h"
  38 #include "ui/browsers/hists.h"
  39 #include "thread.h"
  40 #include "mem2node.h"
  41 #include "symbol.h"
  42 #include "ui/ui.h"
  43 #include "ui/progress.h"
  44 #include "../perf.h"
  45
  46 struct c2c_hists {
  47         struct hists            hists;
  48         struct perf_hpp_list    list;
  49         struct c2c_stats        stats;
  50 };
  51
  52 struct compute_stats {
  53         struct stats             lcl_hitm;
  54         struct stats             rmt_hitm;
  55         struct stats             load;
  56 };
  57
  58 struct c2c_hist_entry {
  59         struct c2c_hists        *hists;
  60         struct c2c_stats         stats;
  61         unsigned long           *cpuset;
  62         unsigned long           *nodeset;
  63         struct c2c_stats        *node_stats;
  64         unsigned int             cacheline_idx;
  65
  66         struct compute_stats     cstats;
  67
  68         unsigned long            paddr;
  69         unsigned long            paddr_cnt;
  70         bool                     paddr_zero;
  71         char                    *nodestr;
  72
  73         /*
  74          * must be at the end,
  75          * because of its callchain dynamic entry
  76          */
  77         struct hist_entry       he;
  78 };
  79
  80 static char const *coalesce_default = "iaddr";
  81
  82 struct perf_c2c {
  83         struct perf_tool        tool;
  84         struct c2c_hists        hists;
  85         struct mem2node         mem2node;
  86
  87         unsigned long           **nodes;
  88         int                      nodes_cnt;
  89         int                      cpus_cnt;
  90         int                     *cpu2node;
  91         int                      node_info;
  92
  93         bool                     show_src;
  94         bool                     show_all;
  95         bool                     use_stdio;
  96         bool                     stats_only;
  97         bool                     symbol_full;
  98         bool                     stitch_lbr;
  99
 100         /* HITM shared clines stats */
 101         struct c2c_stats        hitm_stats;
 102         int                     shared_clines;
 103
 104         int                      display;
 105
 106         const char              *coalesce;
 107         char                    *cl_sort;
 108         char                    *cl_resort;
 109         char                    *cl_output;
 110 };
 111
 112 enum {
 113         DISPLAY_LCL,
 114         DISPLAY_RMT,
 115         DISPLAY_TOT,
 116         DISPLAY_MAX,
 117 };
 118
 119 static const char *display_str[DISPLAY_MAX] = {
 120         [DISPLAY_LCL] = "Local",
 121         [DISPLAY_RMT] = "Remote",
 122         [DISPLAY_TOT] = "Total",
 123 };
 124
 125 static const struct option c2c_options[] = {
 126         OPT_INCR('v', "verbose", &verbose, "be more verbose (show counter open errors, etc)"),
 127         OPT_END()
 128 };
 129
 130 static struct perf_c2c c2c;
 131
 132 static void *c2c_he_zalloc(size_t size)
 133 {
 134         struct c2c_hist_entry *c2c_he;
 135
 136         c2c_he = zalloc(size + sizeof(*c2c_he));
 137         if (!c2c_he)
 138                 return NULL;
 139
 140         c2c_he->cpuset = bitmap_alloc(c2c.cpus_cnt);
 141         if (!c2c_he->cpuset)
 142                 return NULL;
 143
 144         c2c_he->nodeset = bitmap_alloc(c2c.nodes_cnt);
 145         if (!c2c_he->nodeset)
 146                 return NULL;
 147
 148         c2c_he->node_stats = zalloc(c2c.nodes_cnt * sizeof(*c2c_he->node_stats));
 149         if (!c2c_he->node_stats)
 150                 return NULL;
 151
 152         init_stats(&c2c_he->cstats.lcl_hitm);
 153         init_stats(&c2c_he->cstats.rmt_hitm);
 154         init_stats(&c2c_he->cstats.load);
 155
 156         return &c2c_he->he;
 157 }
 158
 159 static void c2c_he_free(void *he)
 160 {
 161         struct c2c_hist_entry *c2c_he;
 162
 163         c2c_he = container_of(he, struct c2c_hist_entry, he);
 164         if (c2c_he->hists) {
 165                 hists__delete_entries(&c2c_he->hists->hists);
 166                 free(c2c_he->hists);
 167         }
 168
 169         free(c2c_he->cpuset);
 170         free(c2c_he->nodeset);
 171         free(c2c_he->nodestr);
 172         free(c2c_he->node_stats);
 173         free(c2c_he);
 174 }
 175
 176 static struct hist_entry_ops c2c_entry_ops = {
 177         .new    = c2c_he_zalloc,
 178         .free   = c2c_he_free,
 179 };
 180
 181 static int c2c_hists__init(struct c2c_hists *hists,
 182                            const char *sort,
 183                            int nr_header_lines);
 184
 185 static struct c2c_hists*
 186 he__get_c2c_hists(struct hist_entry *he,
 187                   const char *sort,
 188                   int nr_header_lines)
 189 {
 190         struct c2c_hist_entry *c2c_he;
 191         struct c2c_hists *hists;
 192         int ret;
 193
 194         c2c_he = container_of(he, struct c2c_hist_entry, he);
 195         if (c2c_he->hists)
 196                 return c2c_he->hists;
 197
 198         hists = c2c_he->hists = zalloc(sizeof(*hists));
 199         if (!hists)
 200                 return NULL;
 201
 202         ret = c2c_hists__init(hists, sort, nr_header_lines);
 203         if (ret) {
 204                 free(hists);
 205                 return NULL;
 206         }
 207
 208         return hists;
 209 }
 210
 211 static void c2c_he__set_cpu(struct c2c_hist_entry *c2c_he,
 212                             struct perf_sample *sample)
 213 {
 214         if (WARN_ONCE(sample->cpu == (unsigned int) -1,
 215                       "WARNING: no sample cpu value"))
 216                 return;
 217
 218         set_bit(sample->cpu, c2c_he->cpuset);
 219 }
 220
 221 static void c2c_he__set_node(struct c2c_hist_entry *c2c_he,
 222                              struct perf_sample *sample)
 223 {
 224         int node;
 225
 226         if (!sample->phys_addr) {
 227                 c2c_he->paddr_zero = true;
 228                 return;
 229         }
 230
 231         node = mem2node__node(&c2c.mem2node, sample->phys_addr);
 232         if (WARN_ONCE(node < 0, "WARNING: failed to find node\n"))
 233                 return;
 234
 235         set_bit(node, c2c_he->nodeset);
 236
 237         if (c2c_he->paddr != sample->phys_addr) {
 238                 c2c_he->paddr_cnt++;
 239                 c2c_he->paddr = sample->phys_addr;
 240         }
 241 }
 242
 243 static void compute_stats(struct c2c_hist_entry *c2c_he,
 244                           struct c2c_stats *stats,
 245                           u64 weight)
 246 {
 247         struct compute_stats *cstats = &c2c_he->cstats;
 248
 249         if (stats->rmt_hitm)
 250                 update_stats(&cstats->rmt_hitm, weight);
 251         else if (stats->lcl_hitm)
 252                 update_stats(&cstats->lcl_hitm, weight);
 253         else if (stats->load)
 254                 update_stats(&cstats->load, weight);
 255 }
 256
 257 static int process_sample_event(struct perf_tool *tool __maybe_unused,
 258                                 union perf_event *event,
 259                                 struct perf_sample *sample,
 260                                 struct evsel *evsel,
 261                                 struct machine *machine)
 262 {
 263         struct c2c_hists *c2c_hists = &c2c.hists;
 264         struct c2c_hist_entry *c2c_he;
 265         struct c2c_stats stats = { .nr_entries = 0, };
 266         struct hist_entry *he;
 267         struct addr_location al;
 268         struct mem_info *mi, *mi_dup;
 269         int ret;
 270
 271         if (machine__resolve(machine, &al, sample) < 0) {
 272                 pr_debug("problem processing %d event, skipping it.\n",
 273                          event->header.type);
 274                 return -1;
 275         }
 276
 277         if (c2c.stitch_lbr)
 278                 al.thread->lbr_stitch_enable = true;
 279
 280         ret = sample__resolve_callchain(sample, &callchain_cursor, NULL,
 281                                         evsel, &al, sysctl_perf_event_max_stack);
 282         if (ret)
 283                 goto out;
 284
 285         mi = sample__resolve_mem(sample, &al);
 286         if (mi == NULL)
 287                 return -ENOMEM;
 288
 289         /*
 290          * The mi object is released in hists__add_entry_ops,
 291          * if it gets sorted out into existing data, so we need
 292          * to take the copy now.
 293          */
 294         mi_dup = mem_info__get(mi);
 295
 296         c2c_decode_stats(&stats, mi);
 297
 298         he = hists__add_entry_ops(&c2c_hists->hists, &c2c_entry_ops,
 299                                   &al, NULL, NULL, mi,
 300                                   sample, true);
 301         if (he == NULL)
 302                 goto free_mi;
 303
 304         c2c_he = container_of(he, struct c2c_hist_entry, he);
 305         c2c_add_stats(&c2c_he->stats, &stats);
 306         c2c_add_stats(&c2c_hists->stats, &stats);
 307
 308         c2c_he__set_cpu(c2c_he, sample);
 309         c2c_he__set_node(c2c_he, sample);
 310
 311         hists__inc_nr_samples(&c2c_hists->hists, he->filtered);
 312         ret = hist_entry__append_callchain(he, sample);
 313
 314         if (!ret) {
 315                 /*
 316                  * There's already been warning about missing
 317                  * sample's cpu value. Let's account all to
 318                  * node 0 in this case, without any further
 319                  * warning.
 320                  *
 321                  * Doing node stats only for single callchain data.
 322                  */
 323                 int cpu = sample->cpu == (unsigned int) -1 ? 0 : sample->cpu;
 324                 int node = c2c.cpu2node[cpu];
 325
 326                 mi = mi_dup;
 327
 328                 c2c_hists = he__get_c2c_hists(he, c2c.cl_sort, 2);
 329                 if (!c2c_hists)
 330                         goto free_mi;
 331
 332                 he = hists__add_entry_ops(&c2c_hists->hists, &c2c_entry_ops,
 333                                           &al, NULL, NULL, mi,
 334                                           sample, true);
 335                 if (he == NULL)
 336                         goto free_mi;
 337
 338                 c2c_he = container_of(he, struct c2c_hist_entry, he);
 339                 c2c_add_stats(&c2c_he->stats, &stats);
 340                 c2c_add_stats(&c2c_hists->stats, &stats);
 341                 c2c_add_stats(&c2c_he->node_stats[node], &stats);
 342
 343                 compute_stats(c2c_he, &stats, sample->weight);
 344
 345                 c2c_he__set_cpu(c2c_he, sample);
 346                 c2c_he__set_node(c2c_he, sample);
 347
 348                 hists__inc_nr_samples(&c2c_hists->hists, he->filtered);
 349                 ret = hist_entry__append_callchain(he, sample);
 350         }
 351
 352 out:
 353         addr_location__put(&al);
 354         return ret;
 355
 356 free_mi:
 357         mem_info__put(mi_dup);
 358         mem_info__put(mi);
 359         ret = -ENOMEM;
 360         goto out;
 361 }
 362
 363 static struct perf_c2c c2c = {
 364         .tool = {
 365                 .sample         = process_sample_event,
 366                 .mmap           = perf_event__process_mmap,
 367                 .mmap2          = perf_event__process_mmap2,
 368                 .comm           = perf_event__process_comm,
 369                 .exit           = perf_event__process_exit,
 370                 .fork           = perf_event__process_fork,
 371                 .lost           = perf_event__process_lost,
 372                 .attr           = perf_event__process_attr,
 373                 .auxtrace_info  = perf_event__process_auxtrace_info,
 374                 .auxtrace       = perf_event__process_auxtrace,
 375                 .auxtrace_error = perf_event__process_auxtrace_error,
 376                 .ordered_events = true,
 377                 .ordering_requires_timestamps = true,
 378         },
 379 };
 380
 381 static const char * const c2c_usage[] = {
 382         "perf c2c {record|report}",
 383         NULL
 384 };
 385
 386 static const char * const __usage_report[] = {
 387         "perf c2c report",
 388         NULL
 389 };
 390
 391 static const char * const *report_c2c_usage = __usage_report;
 392
 393 #define C2C_HEADER_MAX 2
 394
 395 struct c2c_header {
 396         struct {
 397                 const char *text;
 398                 int         span;
 399         } line[C2C_HEADER_MAX];
 400 };
 401
 402 struct c2c_dimension {
 403         struct c2c_header        header;
 404         const char              *name;
 405         int                      width;
 406         struct sort_entry       *se;
 407
 408         int64_t (*cmp)(struct perf_hpp_fmt *fmt,
 409                        struct hist_entry *, struct hist_entry *);
 410         int   (*entry)(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 411                        struct hist_entry *he);
 412         int   (*color)(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 413                        struct hist_entry *he);
 414 };
 415
 416 struct c2c_fmt {
 417         struct perf_hpp_fmt      fmt;
 418         struct c2c_dimension    *dim;
 419 };
 420
 421 #define SYMBOL_WIDTH 30
 422
 423 static struct c2c_dimension dim_symbol;
 424 static struct c2c_dimension dim_srcline;
 425
 426 static int symbol_width(struct hists *hists, struct sort_entry *se)
 427 {
 428         int width = hists__col_len(hists, se->se_width_idx);
 429
 430         if (!c2c.symbol_full)
 431                 width = MIN(width, SYMBOL_WIDTH);
 432
 433         return width;
 434 }
 435
 436 static int c2c_width(struct perf_hpp_fmt *fmt,
 437                      struct perf_hpp *hpp __maybe_unused,
 438                      struct hists *hists)
 439 {
 440         struct c2c_fmt *c2c_fmt;
 441         struct c2c_dimension *dim;
 442
 443         c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
 444         dim = c2c_fmt->dim;
 445
 446         if (dim == &dim_symbol || dim == &dim_srcline)
 447                 return symbol_width(hists, dim->se);
 448
 449         return dim->se ? hists__col_len(hists, dim->se->se_width_idx) :
 450                          c2c_fmt->dim->width;
 451 }
 452
 453 static int c2c_header(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 454                       struct hists *hists, int line, int *span)
 455 {
 456         struct perf_hpp_list *hpp_list = hists->hpp_list;
 457         struct c2c_fmt *c2c_fmt;
 458         struct c2c_dimension *dim;
 459         const char *text = NULL;
 460         int width = c2c_width(fmt, hpp, hists);
 461
 462         c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
 463         dim = c2c_fmt->dim;
 464
 465         if (dim->se) {
 466                 text = dim->header.line[line].text;
 467                 /* Use the last line from sort_entry if not defined. */
 468                 if (!text && (line == hpp_list->nr_header_lines - 1))
 469                         text = dim->se->se_header;
 470         } else {
 471                 text = dim->header.line[line].text;
 472
 473                 if (*span) {
 474                         (*span)--;
 475                         return 0;
 476                 } else {
 477                         *span = dim->header.line[line].span;
 478                 }
 479         }
 480
 481         if (text == NULL)
 482                 text = "";
 483
 484         return scnprintf(hpp->buf, hpp->size, "%*s", width, text);
 485 }
 486
 487 #define HEX_STR(__s, __v)                               \
 488 ({                                                      \
 489         scnprintf(__s, sizeof(__s), "0x%" PRIx64, __v); \
 490         __s;                                            \
 491 })
 492
 493 static int64_t
 494 dcacheline_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 495                struct hist_entry *left, struct hist_entry *right)
 496 {
 497         return sort__dcacheline_cmp(left, right);
 498 }
 499
 500 static int dcacheline_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 501                             struct hist_entry *he)
 502 {
 503         uint64_t addr = 0;
 504         int width = c2c_width(fmt, hpp, he->hists);
 505         char buf[20];
 506
 507         if (he->mem_info)
 508                 addr = cl_address(he->mem_info->daddr.addr);
 509
 510         return scnprintf(hpp->buf, hpp->size, "%*s", width, HEX_STR(buf, addr));
 511 }
 512
 513 static int
 514 dcacheline_node_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 515                       struct hist_entry *he)
 516 {
 517         struct c2c_hist_entry *c2c_he;
 518         int width = c2c_width(fmt, hpp, he->hists);
 519
 520         c2c_he = container_of(he, struct c2c_hist_entry, he);
 521         if (WARN_ON_ONCE(!c2c_he->nodestr))
 522                 return 0;
 523
 524         return scnprintf(hpp->buf, hpp->size, "%*s", width, c2c_he->nodestr);
 525 }
 526
 527 static int
 528 dcacheline_node_count(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 529                       struct hist_entry *he)
 530 {
 531         struct c2c_hist_entry *c2c_he;
 532         int width = c2c_width(fmt, hpp, he->hists);
 533
 534         c2c_he = container_of(he, struct c2c_hist_entry, he);
 535         return scnprintf(hpp->buf, hpp->size, "%*lu", width, c2c_he->paddr_cnt);
 536 }
 537
 538 static int offset_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 539                         struct hist_entry *he)
 540 {
 541         uint64_t addr = 0;
 542         int width = c2c_width(fmt, hpp, he->hists);
 543         char buf[20];
 544
 545         if (he->mem_info)
 546                 addr = cl_offset(he->mem_info->daddr.al_addr);
 547
 548         return scnprintf(hpp->buf, hpp->size, "%*s", width, HEX_STR(buf, addr));
 549 }
 550
 551 static int64_t
 552 offset_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 553            struct hist_entry *left, struct hist_entry *right)
 554 {
 555         uint64_t l = 0, r = 0;
 556
 557         if (left->mem_info)
 558                 l = cl_offset(left->mem_info->daddr.addr);
 559         if (right->mem_info)
 560                 r = cl_offset(right->mem_info->daddr.addr);
 561
 562         return (int64_t)(r - l);
 563 }
 564
 565 static int
 566 iaddr_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 567             struct hist_entry *he)
 568 {
 569         uint64_t addr = 0;
 570         int width = c2c_width(fmt, hpp, he->hists);
 571         char buf[20];
 572
 573         if (he->mem_info)
 574                 addr = he->mem_info->iaddr.addr;
 575
 576         return scnprintf(hpp->buf, hpp->size, "%*s", width, HEX_STR(buf, addr));
 577 }
 578
 579 static int64_t
 580 iaddr_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 581           struct hist_entry *left, struct hist_entry *right)
 582 {
 583         return sort__iaddr_cmp(left, right);
 584 }
 585
 586 static int
 587 tot_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 588                struct hist_entry *he)
 589 {
 590         struct c2c_hist_entry *c2c_he;
 591         int width = c2c_width(fmt, hpp, he->hists);
 592         unsigned int tot_hitm;
 593
 594         c2c_he = container_of(he, struct c2c_hist_entry, he);
 595         tot_hitm = c2c_he->stats.lcl_hitm + c2c_he->stats.rmt_hitm;
 596
 597         return scnprintf(hpp->buf, hpp->size, "%*u", width, tot_hitm);
 598 }
 599
 600 static int64_t
 601 tot_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 602              struct hist_entry *left, struct hist_entry *right)
 603 {
 604         struct c2c_hist_entry *c2c_left;
 605         struct c2c_hist_entry *c2c_right;
 606         uint64_t tot_hitm_left;
 607         uint64_t tot_hitm_right;
 608
 609         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 610         c2c_right = container_of(right, struct c2c_hist_entry, he);
 611
 612         tot_hitm_left  = c2c_left->stats.lcl_hitm + c2c_left->stats.rmt_hitm;
 613         tot_hitm_right = c2c_right->stats.lcl_hitm + c2c_right->stats.rmt_hitm;
 614
 615         return tot_hitm_left - tot_hitm_right;
 616 }
 617
 618 #define STAT_FN_ENTRY(__f)                                      \
 619 static int                                                      \
 620 __f ## _entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,   \
 621               struct hist_entry *he)                            \
 622 {                                                               \
 623         struct c2c_hist_entry *c2c_he;                          \
 624         int width = c2c_width(fmt, hpp, he->hists);             \
 625                                                                 \
 626         c2c_he = container_of(he, struct c2c_hist_entry, he);   \
 627         return scnprintf(hpp->buf, hpp->size, "%*u", width,     \
 628                          c2c_he->stats.__f);                    \
 629 }
 630
 631 #define STAT_FN_CMP(__f)                                                \
 632 static int64_t                                                          \
 633 __f ## _cmp(struct perf_hpp_fmt *fmt __maybe_unused,                    \
 634             struct hist_entry *left, struct hist_entry *right)          \
 635 {                                                                       \
 636         struct c2c_hist_entry *c2c_left, *c2c_right;                    \
 637                                                                         \
 638         c2c_left  = container_of(left, struct c2c_hist_entry, he);      \
 639         c2c_right = container_of(right, struct c2c_hist_entry, he);     \
 640         return (uint64_t) c2c_left->stats.__f -                         \
 641                (uint64_t) c2c_right->stats.__f;                         \
 642 }
 643
 644 #define STAT_FN(__f)            \
 645         STAT_FN_ENTRY(__f)      \
 646         STAT_FN_CMP(__f)
 647
 648 STAT_FN(rmt_hitm)
 649 STAT_FN(lcl_hitm)
 650 STAT_FN(store)
 651 STAT_FN(st_l1hit)
 652 STAT_FN(st_l1miss)
 653 STAT_FN(ld_fbhit)
 654 STAT_FN(ld_l1hit)
 655 STAT_FN(ld_l2hit)
 656 STAT_FN(ld_llchit)
 657 STAT_FN(rmt_hit)
 658
 659 static uint64_t total_records(struct c2c_stats *stats)
 660 {
 661         uint64_t lclmiss, ldcnt, total;
 662
 663         lclmiss  = stats->lcl_dram +
 664                    stats->rmt_dram +
 665                    stats->rmt_hitm +
 666                    stats->rmt_hit;
 667
 668         ldcnt    = lclmiss +
 669                    stats->ld_fbhit +
 670                    stats->ld_l1hit +
 671                    stats->ld_l2hit +
 672                    stats->ld_llchit +
 673                    stats->lcl_hitm;
 674
 675         total    = ldcnt +
 676                    stats->st_l1hit +
 677                    stats->st_l1miss;
 678
 679         return total;
 680 }
 681
 682 static int
 683 tot_recs_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 684                 struct hist_entry *he)
 685 {
 686         struct c2c_hist_entry *c2c_he;
 687         int width = c2c_width(fmt, hpp, he->hists);
 688         uint64_t tot_recs;
 689
 690         c2c_he = container_of(he, struct c2c_hist_entry, he);
 691         tot_recs = total_records(&c2c_he->stats);
 692
 693         return scnprintf(hpp->buf, hpp->size, "%*" PRIu64, width, tot_recs);
 694 }
 695
 696 static int64_t
 697 tot_recs_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 698              struct hist_entry *left, struct hist_entry *right)
 699 {
 700         struct c2c_hist_entry *c2c_left;
 701         struct c2c_hist_entry *c2c_right;
 702         uint64_t tot_recs_left;
 703         uint64_t tot_recs_right;
 704
 705         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 706         c2c_right = container_of(right, struct c2c_hist_entry, he);
 707
 708         tot_recs_left  = total_records(&c2c_left->stats);
 709         tot_recs_right = total_records(&c2c_right->stats);
 710
 711         return tot_recs_left - tot_recs_right;
 712 }
 713
 714 static uint64_t total_loads(struct c2c_stats *stats)
 715 {
 716         uint64_t lclmiss, ldcnt;
 717
 718         lclmiss  = stats->lcl_dram +
 719                    stats->rmt_dram +
 720                    stats->rmt_hitm +
 721                    stats->rmt_hit;
 722
 723         ldcnt    = lclmiss +
 724                    stats->ld_fbhit +
 725                    stats->ld_l1hit +
 726                    stats->ld_l2hit +
 727                    stats->ld_llchit +
 728                    stats->lcl_hitm;
 729
 730         return ldcnt;
 731 }
 732
 733 static int
 734 tot_loads_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 735                 struct hist_entry *he)
 736 {
 737         struct c2c_hist_entry *c2c_he;
 738         int width = c2c_width(fmt, hpp, he->hists);
 739         uint64_t tot_recs;
 740
 741         c2c_he = container_of(he, struct c2c_hist_entry, he);
 742         tot_recs = total_loads(&c2c_he->stats);
 743
 744         return scnprintf(hpp->buf, hpp->size, "%*" PRIu64, width, tot_recs);
 745 }
 746
 747 static int64_t
 748 tot_loads_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 749               struct hist_entry *left, struct hist_entry *right)
 750 {
 751         struct c2c_hist_entry *c2c_left;
 752         struct c2c_hist_entry *c2c_right;
 753         uint64_t tot_recs_left;
 754         uint64_t tot_recs_right;
 755
 756         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 757         c2c_right = container_of(right, struct c2c_hist_entry, he);
 758
 759         tot_recs_left  = total_loads(&c2c_left->stats);
 760         tot_recs_right = total_loads(&c2c_right->stats);
 761
 762         return tot_recs_left - tot_recs_right;
 763 }
 764
 765 typedef double (get_percent_cb)(struct c2c_hist_entry *);
 766
 767 static int
 768 percent_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 769               struct hist_entry *he, get_percent_cb get_percent)
 770 {
 771         struct c2c_hist_entry *c2c_he;
 772         int width = c2c_width(fmt, hpp, he->hists);
 773         double per;
 774
 775         c2c_he = container_of(he, struct c2c_hist_entry, he);
 776         per = get_percent(c2c_he);
 777
 778 #ifdef HAVE_SLANG_SUPPORT
 779         if (use_browser)
 780                 return __hpp__slsmg_color_printf(hpp, "%*.2f%%", width - 1, per);
 781 #endif
 782         return hpp_color_scnprintf(hpp, "%*.2f%%", width - 1, per);
 783 }
 784
 785 static double percent_hitm(struct c2c_hist_entry *c2c_he)
 786 {
 787         struct c2c_hists *hists;
 788         struct c2c_stats *stats;
 789         struct c2c_stats *total;
 790         int tot = 0, st = 0;
 791         double p;
 792
 793         hists = container_of(c2c_he->he.hists, struct c2c_hists, hists);
 794         stats = &c2c_he->stats;
 795         total = &hists->stats;
 796
 797         switch (c2c.display) {
 798         case DISPLAY_RMT:
 799                 st  = stats->rmt_hitm;
 800                 tot = total->rmt_hitm;
 801                 break;
 802         case DISPLAY_LCL:
 803                 st  = stats->lcl_hitm;
 804                 tot = total->lcl_hitm;
 805                 break;
 806         case DISPLAY_TOT:
 807                 st  = stats->tot_hitm;
 808                 tot = total->tot_hitm;
 809         default:
 810                 break;
 811         }
 812
 813         p = tot ? (double) st / tot : 0;
 814
 815         return 100 * p;
 816 }
 817
 818 #define PERC_STR(__s, __v)                              \
 819 ({                                                      \
 820         scnprintf(__s, sizeof(__s), "%.2F%%", __v);     \
 821         __s;                                            \
 822 })
 823
 824 static int
 825 percent_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 826                    struct hist_entry *he)
 827 {
 828         struct c2c_hist_entry *c2c_he;
 829         int width = c2c_width(fmt, hpp, he->hists);
 830         char buf[10];
 831         double per;
 832
 833         c2c_he = container_of(he, struct c2c_hist_entry, he);
 834         per = percent_hitm(c2c_he);
 835         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 836 }
 837
 838 static int
 839 percent_hitm_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 840                    struct hist_entry *he)
 841 {
 842         return percent_color(fmt, hpp, he, percent_hitm);
 843 }
 844
 845 static int64_t
 846 percent_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 847                  struct hist_entry *left, struct hist_entry *right)
 848 {
 849         struct c2c_hist_entry *c2c_left;
 850         struct c2c_hist_entry *c2c_right;
 851         double per_left;
 852         double per_right;
 853
 854         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 855         c2c_right = container_of(right, struct c2c_hist_entry, he);
 856
 857         per_left  = percent_hitm(c2c_left);
 858         per_right = percent_hitm(c2c_right);
 859
 860         return per_left - per_right;
 861 }
 862
 863 static struct c2c_stats *he_stats(struct hist_entry *he)
 864 {
 865         struct c2c_hist_entry *c2c_he;
 866
 867         c2c_he = container_of(he, struct c2c_hist_entry, he);
 868         return &c2c_he->stats;
 869 }
 870
 871 static struct c2c_stats *total_stats(struct hist_entry *he)
 872 {
 873         struct c2c_hists *hists;
 874
 875         hists = container_of(he->hists, struct c2c_hists, hists);
 876         return &hists->stats;
 877 }
 878
 879 static double percent(int st, int tot)
 880 {
 881         return tot ? 100. * (double) st / (double) tot : 0;
 882 }
 883
 884 #define PERCENT(__h, __f) percent(he_stats(__h)->__f, total_stats(__h)->__f)
 885
 886 #define PERCENT_FN(__f)                                                         \
 887 static double percent_ ## __f(struct c2c_hist_entry *c2c_he)                    \
 888 {                                                                               \
 889         struct c2c_hists *hists;                                                \
 890                                                                                 \
 891         hists = container_of(c2c_he->he.hists, struct c2c_hists, hists);        \
 892         return percent(c2c_he->stats.__f, hists->stats.__f);                    \
 893 }
 894
 895 PERCENT_FN(rmt_hitm)
 896 PERCENT_FN(lcl_hitm)
 897 PERCENT_FN(st_l1hit)
 898 PERCENT_FN(st_l1miss)
 899
 900 static int
 901 percent_rmt_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 902                        struct hist_entry *he)
 903 {
 904         int width = c2c_width(fmt, hpp, he->hists);
 905         double per = PERCENT(he, rmt_hitm);
 906         char buf[10];
 907
 908         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 909 }
 910
 911 static int
 912 percent_rmt_hitm_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 913                        struct hist_entry *he)
 914 {
 915         return percent_color(fmt, hpp, he, percent_rmt_hitm);
 916 }
 917
 918 static int64_t
 919 percent_rmt_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 920                      struct hist_entry *left, struct hist_entry *right)
 921 {
 922         double per_left;
 923         double per_right;
 924
 925         per_left  = PERCENT(left, lcl_hitm);
 926         per_right = PERCENT(right, lcl_hitm);
 927
 928         return per_left - per_right;
 929 }
 930
 931 static int
 932 percent_lcl_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 933                        struct hist_entry *he)
 934 {
 935         int width = c2c_width(fmt, hpp, he->hists);
 936         double per = PERCENT(he, lcl_hitm);
 937         char buf[10];
 938
 939         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 940 }
 941
 942 static int
 943 percent_lcl_hitm_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 944                        struct hist_entry *he)
 945 {
 946         return percent_color(fmt, hpp, he, percent_lcl_hitm);
 947 }
 948
 949 static int64_t
 950 percent_lcl_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 951                      struct hist_entry *left, struct hist_entry *right)
 952 {
 953         double per_left;
 954         double per_right;
 955
 956         per_left  = PERCENT(left, lcl_hitm);
 957         per_right = PERCENT(right, lcl_hitm);
 958
 959         return per_left - per_right;
 960 }
 961
 962 static int
 963 percent_stores_l1hit_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 964                            struct hist_entry *he)
 965 {
 966         int width = c2c_width(fmt, hpp, he->hists);
 967         double per = PERCENT(he, st_l1hit);
 968         char buf[10];
 969
 970         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 971 }
 972
 973 static int
 974 percent_stores_l1hit_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 975                            struct hist_entry *he)
 976 {
 977         return percent_color(fmt, hpp, he, percent_st_l1hit);
 978 }
 979
 980 static int64_t
 981 percent_stores_l1hit_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 982                         struct hist_entry *left, struct hist_entry *right)
 983 {
 984         double per_left;
 985         double per_right;
 986
 987         per_left  = PERCENT(left, st_l1hit);
 988         per_right = PERCENT(right, st_l1hit);
 989
 990         return per_left - per_right;
 991 }
 992
 993 static int
 994 percent_stores_l1miss_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 995                            struct hist_entry *he)
 996 {
 997         int width = c2c_width(fmt, hpp, he->hists);
 998         double per = PERCENT(he, st_l1miss);
 999         char buf[10];
1000
1001         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
1002 }
1003
1004 static int
1005 percent_stores_l1miss_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1006                             struct hist_entry *he)
1007 {
1008         return percent_color(fmt, hpp, he, percent_st_l1miss);
1009 }
1010
1011 static int64_t
1012 percent_stores_l1miss_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
1013                           struct hist_entry *left, struct hist_entry *right)
1014 {
1015         double per_left;
1016         double per_right;
1017
1018         per_left  = PERCENT(left, st_l1miss);
1019         per_right = PERCENT(right, st_l1miss);
1020
1021         return per_left - per_right;
1022 }
1023
1024 STAT_FN(lcl_dram)
1025 STAT_FN(rmt_dram)
1026
1027 static int
1028 pid_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1029           struct hist_entry *he)
1030 {
1031         int width = c2c_width(fmt, hpp, he->hists);
1032
1033         return scnprintf(hpp->buf, hpp->size, "%*d", width, he->thread->pid_);
1034 }
1035
1036 static int64_t
1037 pid_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
1038         struct hist_entry *left, struct hist_entry *right)
1039 {
1040         return left->thread->pid_ - right->thread->pid_;
1041 }
1042
1043 static int64_t
1044 empty_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
1045           struct hist_entry *left __maybe_unused,
1046           struct hist_entry *right __maybe_unused)
1047 {
1048         return 0;
1049 }
1050
1051 static int
1052 node_entry(struct perf_hpp_fmt *fmt __maybe_unused, struct perf_hpp *hpp,
1053            struct hist_entry *he)
1054 {
1055         struct c2c_hist_entry *c2c_he;
1056         bool first = true;
1057         int node;
1058         int ret = 0;
1059
1060         c2c_he = container_of(he, struct c2c_hist_entry, he);
1061
1062         for (node = 0; node < c2c.nodes_cnt; node++) {
1063                 DECLARE_BITMAP(set, c2c.cpus_cnt);
1064
1065                 bitmap_zero(set, c2c.cpus_cnt);
1066                 bitmap_and(set, c2c_he->cpuset, c2c.nodes[node], c2c.cpus_cnt);
1067
1068                 if (!bitmap_weight(set, c2c.cpus_cnt)) {
1069                         if (c2c.node_info == 1) {
1070                                 ret = scnprintf(hpp->buf, hpp->size, "%21s", " ");
1071                                 advance_hpp(hpp, ret);
1072                         }
1073                         continue;
1074                 }
1075
1076                 if (!first) {
1077                         ret = scnprintf(hpp->buf, hpp->size, " ");
1078                         advance_hpp(hpp, ret);
1079                 }
1080
1081                 switch (c2c.node_info) {
1082                 case 0:
1083                         ret = scnprintf(hpp->buf, hpp->size, "%2d", node);
1084                         advance_hpp(hpp, ret);
1085                         break;
1086                 case 1:
1087                 {
1088                         int num = bitmap_weight(set, c2c.cpus_cnt);
1089                         struct c2c_stats *stats = &c2c_he->node_stats[node];
1090
1091                         ret = scnprintf(hpp->buf, hpp->size, "%2d{%2d ", node, num);
1092                         advance_hpp(hpp, ret);
1093
1094                 #define DISPLAY_HITM(__h)                                               \
1095                         if (c2c_he->stats.__h> 0) {                                     \
1096                                 ret = scnprintf(hpp->buf, hpp->size, "%5.1f%% ",        \
1097                                                 percent(stats->__h, c2c_he->stats.__h));\
1098                         } else {                                                        \
1099                                 ret = scnprintf(hpp->buf, hpp->size, "%6s ", "n/a");    \
1100                         }
1101
1102                         switch (c2c.display) {
1103                         case DISPLAY_RMT:
1104                                 DISPLAY_HITM(rmt_hitm);
1105                                 break;
1106                         case DISPLAY_LCL:
1107                                 DISPLAY_HITM(lcl_hitm);
1108                                 break;
1109                         case DISPLAY_TOT:
1110                                 DISPLAY_HITM(tot_hitm);
1111                         default:
1112                                 break;
1113                         }
1114
1115                 #undef DISPLAY_HITM
1116
1117                         advance_hpp(hpp, ret);
1118
1119                         if (c2c_he->stats.store > 0) {
1120                                 ret = scnprintf(hpp->buf, hpp->size, "%5.1f%%}",
1121                                                 percent(stats->store, c2c_he->stats.store));
1122                         } else {
1123                                 ret = scnprintf(hpp->buf, hpp->size, "%6s}", "n/a");
1124                         }
1125
1126                         advance_hpp(hpp, ret);
1127                         break;
1128                 }
1129                 case 2:
1130                         ret = scnprintf(hpp->buf, hpp->size, "%2d{", node);
1131                         advance_hpp(hpp, ret);
1132
1133                         ret = bitmap_scnprintf(set, c2c.cpus_cnt, hpp->buf, hpp->size);
1134                         advance_hpp(hpp, ret);
1135
1136                         ret = scnprintf(hpp->buf, hpp->size, "}");
1137                         advance_hpp(hpp, ret);
1138                         break;
1139                 default:
1140                         break;
1141                 }
1142
1143                 first = false;
1144         }
1145
1146         return 0;
1147 }
1148
1149 static int
1150 mean_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1151            struct hist_entry *he, double mean)
1152 {
1153         int width = c2c_width(fmt, hpp, he->hists);
1154         char buf[10];
1155
1156         scnprintf(buf, 10, "%6.0f", mean);
1157         return scnprintf(hpp->buf, hpp->size, "%*s", width, buf);
1158 }
1159
1160 #define MEAN_ENTRY(__func, __val)                                               \
1161 static int                                                                      \
1162 __func(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, struct hist_entry *he)   \
1163 {                                                                               \
1164         struct c2c_hist_entry *c2c_he;                                          \
1165         c2c_he = container_of(he, struct c2c_hist_entry, he);                   \
1166         return mean_entry(fmt, hpp, he, avg_stats(&c2c_he->cstats.__val));      \
1167 }
1168
1169 MEAN_ENTRY(mean_rmt_entry,  rmt_hitm);
1170 MEAN_ENTRY(mean_lcl_entry,  lcl_hitm);
1171 MEAN_ENTRY(mean_load_entry, load);
1172
1173 static int
1174 cpucnt_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1175              struct hist_entry *he)
1176 {
1177         struct c2c_hist_entry *c2c_he;
1178         int width = c2c_width(fmt, hpp, he->hists);
1179         char buf[10];
1180
1181         c2c_he = container_of(he, struct c2c_hist_entry, he);
1182
1183         scnprintf(buf, 10, "%d", bitmap_weight(c2c_he->cpuset, c2c.cpus_cnt));
1184         return scnprintf(hpp->buf, hpp->size, "%*s", width, buf);
1185 }
1186
1187 static int
1188 cl_idx_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1189              struct hist_entry *he)
1190 {
1191         struct c2c_hist_entry *c2c_he;
1192         int width = c2c_width(fmt, hpp, he->hists);
1193         char buf[10];
1194
1195         c2c_he = container_of(he, struct c2c_hist_entry, he);
1196
1197         scnprintf(buf, 10, "%u", c2c_he->cacheline_idx);
1198         return scnprintf(hpp->buf, hpp->size, "%*s", width, buf);
1199 }
1200
1201 static int
1202 cl_idx_empty_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1203                    struct hist_entry *he)
1204 {
1205         int width = c2c_width(fmt, hpp, he->hists);
1206
1207         return scnprintf(hpp->buf, hpp->size, "%*s", width, "");
1208 }
1209
1210 #define HEADER_LOW(__h)                 \
1211         {                               \
1212                 .line[1] = {            \
1213                         .text = __h,    \
1214                 },                      \
1215         }
1216
1217 #define HEADER_BOTH(__h0, __h1)         \
1218         {                               \
1219                 .line[0] = {            \
1220                         .text = __h0,   \
1221                 },                      \
1222                 .line[1] = {            \
1223                         .text = __h1,   \
1224                 },                      \
1225         }
1226
1227 #define HEADER_SPAN(__h0, __h1, __s)    \
1228         {                               \
1229                 .line[0] = {            \
1230                         .text = __h0,   \
1231                         .span = __s,    \
1232                 },                      \
1233                 .line[1] = {            \
1234                         .text = __h1,   \
1235                 },                      \
1236         }
1237
1238 #define HEADER_SPAN_LOW(__h)            \
1239         {                               \
1240                 .line[1] = {            \
1241                         .text = __h,    \
1242                 },                      \
1243         }
1244
1245 static struct c2c_dimension dim_dcacheline = {
1246         .header         = HEADER_SPAN("--- Cacheline ----", "Address", 2),
1247         .name           = "dcacheline",
1248         .cmp            = dcacheline_cmp,
1249         .entry          = dcacheline_entry,
1250         .width          = 18,
1251 };
1252
1253 static struct c2c_dimension dim_dcacheline_node = {
1254         .header         = HEADER_LOW("Node"),
1255         .name           = "dcacheline_node",
1256         .cmp            = empty_cmp,
1257         .entry          = dcacheline_node_entry,
1258         .width          = 4,
1259 };
1260
1261 static struct c2c_dimension dim_dcacheline_count = {
1262         .header         = HEADER_LOW("PA cnt"),
1263         .name           = "dcacheline_count",
1264         .cmp            = empty_cmp,
1265         .entry          = dcacheline_node_count,
1266         .width          = 6,
1267 };
1268
1269 static struct c2c_header header_offset_tui = HEADER_SPAN("-----", "Off", 2);
1270
1271 static struct c2c_dimension dim_offset = {
1272         .header         = HEADER_SPAN("--- Data address -", "Offset", 2),
1273         .name           = "offset",
1274         .cmp            = offset_cmp,
1275         .entry          = offset_entry,
1276         .width          = 18,
1277 };
1278
1279 static struct c2c_dimension dim_offset_node = {
1280         .header         = HEADER_LOW("Node"),
1281         .name           = "offset_node",
1282         .cmp            = empty_cmp,
1283         .entry          = dcacheline_node_entry,
1284         .width          = 4,
1285 };
1286
1287 static struct c2c_dimension dim_iaddr = {
1288         .header         = HEADER_LOW("Code address"),
1289         .name           = "iaddr",
1290         .cmp            = iaddr_cmp,
1291         .entry          = iaddr_entry,
1292         .width          = 18,
1293 };
1294
1295 static struct c2c_dimension dim_tot_hitm = {
1296         .header         = HEADER_SPAN("------- Load Hitm -------", "Total", 2),
1297         .name           = "tot_hitm",
1298         .cmp            = tot_hitm_cmp,
1299         .entry          = tot_hitm_entry,
1300         .width          = 7,
1301 };
1302
1303 static struct c2c_dimension dim_lcl_hitm = {
1304         .header         = HEADER_SPAN_LOW("LclHitm"),
1305         .name           = "lcl_hitm",
1306         .cmp            = lcl_hitm_cmp,
1307         .entry          = lcl_hitm_entry,
1308         .width          = 7,
1309 };
1310
1311 static struct c2c_dimension dim_rmt_hitm = {
1312         .header         = HEADER_SPAN_LOW("RmtHitm"),
1313         .name           = "rmt_hitm",
1314         .cmp            = rmt_hitm_cmp,
1315         .entry          = rmt_hitm_entry,
1316         .width          = 7,
1317 };
1318
1319 static struct c2c_dimension dim_cl_rmt_hitm = {
1320         .header         = HEADER_SPAN("----- HITM -----", "Rmt", 1),
1321         .name           = "cl_rmt_hitm",
1322         .cmp            = rmt_hitm_cmp,
1323         .entry          = rmt_hitm_entry,
1324         .width          = 7,
1325 };
1326
1327 static struct c2c_dimension dim_cl_lcl_hitm = {
1328         .header         = HEADER_SPAN_LOW("Lcl"),
1329         .name           = "cl_lcl_hitm",
1330         .cmp            = lcl_hitm_cmp,
1331         .entry          = lcl_hitm_entry,
1332         .width          = 7,
1333 };
1334
1335 static struct c2c_dimension dim_tot_stores = {
1336         .header         = HEADER_BOTH("Total", "Stores"),
1337         .name           = "tot_stores",
1338         .cmp            = store_cmp,
1339         .entry          = store_entry,
1340         .width          = 7,
1341 };
1342
1343 static struct c2c_dimension dim_stores_l1hit = {
1344         .header         = HEADER_SPAN("---- Stores ----", "L1Hit", 1),
1345         .name           = "stores_l1hit",
1346         .cmp            = st_l1hit_cmp,
1347         .entry          = st_l1hit_entry,
1348         .width          = 7,
1349 };
1350
1351 static struct c2c_dimension dim_stores_l1miss = {
1352         .header         = HEADER_SPAN_LOW("L1Miss"),
1353         .name           = "stores_l1miss",
1354         .cmp            = st_l1miss_cmp,
1355         .entry          = st_l1miss_entry,
1356         .width          = 7,
1357 };
1358
1359 static struct c2c_dimension dim_cl_stores_l1hit = {
1360         .header         = HEADER_SPAN("-- Store Refs --", "L1 Hit", 1),
1361         .name           = "cl_stores_l1hit",
1362         .cmp            = st_l1hit_cmp,
1363         .entry          = st_l1hit_entry,
1364         .width          = 7,
1365 };
1366
1367 static struct c2c_dimension dim_cl_stores_l1miss = {
1368         .header         = HEADER_SPAN_LOW("L1 Miss"),
1369         .name           = "cl_stores_l1miss",
1370         .cmp            = st_l1miss_cmp,
1371         .entry          = st_l1miss_entry,
1372         .width          = 7,
1373 };
1374
1375 static struct c2c_dimension dim_ld_fbhit = {
1376         .header         = HEADER_SPAN("----- Core Load Hit -----", "FB", 2),
1377         .name           = "ld_fbhit",
1378         .cmp            = ld_fbhit_cmp,
1379         .entry          = ld_fbhit_entry,
1380         .width          = 7,
1381 };
1382
1383 static struct c2c_dimension dim_ld_l1hit = {
1384         .header         = HEADER_SPAN_LOW("L1"),
1385         .name           = "ld_l1hit",
1386         .cmp            = ld_l1hit_cmp,
1387         .entry          = ld_l1hit_entry,
1388         .width          = 7,
1389 };
1390
1391 static struct c2c_dimension dim_ld_l2hit = {
1392         .header         = HEADER_SPAN_LOW("L2"),
1393         .name           = "ld_l2hit",
1394         .cmp            = ld_l2hit_cmp,
1395         .entry          = ld_l2hit_entry,
1396         .width          = 7,
1397 };
1398
1399 static struct c2c_dimension dim_ld_llchit = {
1400         .header         = HEADER_SPAN("- LLC Load Hit --", "LclHit", 1),
1401         .name           = "ld_lclhit",
1402         .cmp            = ld_llchit_cmp,
1403         .entry          = ld_llchit_entry,
1404         .width          = 8,
1405 };
1406
1407 static struct c2c_dimension dim_ld_rmthit = {
1408         .header         = HEADER_SPAN("- RMT Load Hit --", "RmtHit", 1),
1409         .name           = "ld_rmthit",
1410         .cmp            = rmt_hit_cmp,
1411         .entry          = rmt_hit_entry,
1412         .width          = 8,
1413 };
1414
1415 static struct c2c_dimension dim_tot_recs = {
1416         .header         = HEADER_BOTH("Total", "records"),
1417         .name           = "tot_recs",
1418         .cmp            = tot_recs_cmp,
1419         .entry          = tot_recs_entry,
1420         .width          = 7,
1421 };
1422
1423 static struct c2c_dimension dim_tot_loads = {
1424         .header         = HEADER_BOTH("Total", "Loads"),
1425         .name           = "tot_loads",
1426         .cmp            = tot_loads_cmp,
1427         .entry          = tot_loads_entry,
1428         .width          = 7,
1429 };
1430
1431 static struct c2c_header percent_hitm_header[] = {
1432         [DISPLAY_LCL] = HEADER_BOTH("Lcl", "Hitm"),
1433         [DISPLAY_RMT] = HEADER_BOTH("Rmt", "Hitm"),
1434         [DISPLAY_TOT] = HEADER_BOTH("Tot", "Hitm"),
1435 };
1436
1437 static struct c2c_dimension dim_percent_hitm = {
1438         .name           = "percent_hitm",
1439         .cmp            = percent_hitm_cmp,
1440         .entry          = percent_hitm_entry,
1441         .color          = percent_hitm_color,
1442         .width          = 7,
1443 };
1444
1445 static struct c2c_dimension dim_percent_rmt_hitm = {
1446         .header         = HEADER_SPAN("----- HITM -----", "RmtHitm", 1),
1447         .name           = "percent_rmt_hitm",
1448         .cmp            = percent_rmt_hitm_cmp,
1449         .entry          = percent_rmt_hitm_entry,
1450         .color          = percent_rmt_hitm_color,
1451         .width          = 7,
1452 };
1453
1454 static struct c2c_dimension dim_percent_lcl_hitm = {
1455         .header         = HEADER_SPAN_LOW("LclHitm"),
1456         .name           = "percent_lcl_hitm",
1457         .cmp            = percent_lcl_hitm_cmp,
1458         .entry          = percent_lcl_hitm_entry,
1459         .color          = percent_lcl_hitm_color,
1460         .width          = 7,
1461 };
1462
1463 static struct c2c_dimension dim_percent_stores_l1hit = {
1464         .header         = HEADER_SPAN("-- Store Refs --", "L1 Hit", 1),
1465         .name           = "percent_stores_l1hit",
1466         .cmp            = percent_stores_l1hit_cmp,
1467         .entry          = percent_stores_l1hit_entry,
1468         .color          = percent_stores_l1hit_color,
1469         .width          = 7,
1470 };
1471
1472 static struct c2c_dimension dim_percent_stores_l1miss = {
1473         .header         = HEADER_SPAN_LOW("L1 Miss"),
1474         .name           = "percent_stores_l1miss",
1475         .cmp            = percent_stores_l1miss_cmp,
1476         .entry          = percent_stores_l1miss_entry,
1477         .color          = percent_stores_l1miss_color,
1478         .width          = 7,
1479 };
1480
1481 static struct c2c_dimension dim_dram_lcl = {
1482         .header         = HEADER_SPAN("--- Load Dram ----", "Lcl", 1),
1483         .name           = "dram_lcl",
1484         .cmp            = lcl_dram_cmp,
1485         .entry          = lcl_dram_entry,
1486         .width          = 8,
1487 };
1488
1489 static struct c2c_dimension dim_dram_rmt = {
1490         .header         = HEADER_SPAN_LOW("Rmt"),
1491         .name           = "dram_rmt",
1492         .cmp            = rmt_dram_cmp,
1493         .entry          = rmt_dram_entry,
1494         .width          = 8,
1495 };
1496
1497 static struct c2c_dimension dim_pid = {
1498         .header         = HEADER_LOW("Pid"),
1499         .name           = "pid",
1500         .cmp            = pid_cmp,
1501         .entry          = pid_entry,
1502         .width          = 7,
1503 };
1504
1505 static struct c2c_dimension dim_tid = {
1506         .header         = HEADER_LOW("Tid"),
1507         .name           = "tid",
1508         .se             = &sort_thread,
1509 };
1510
1511 static struct c2c_dimension dim_symbol = {
1512         .name           = "symbol",
1513         .se             = &sort_sym,
1514 };
1515
1516 static struct c2c_dimension dim_dso = {
1517         .header         = HEADER_BOTH("Shared", "Object"),
1518         .name           = "dso",
1519         .se             = &sort_dso,
1520 };
1521
1522 static struct c2c_header header_node[3] = {
1523         HEADER_LOW("Node"),
1524         HEADER_LOW("Node{cpus %hitms %stores}"),
1525         HEADER_LOW("Node{cpu list}"),
1526 };
1527
1528 static struct c2c_dimension dim_node = {
1529         .name           = "node",
1530         .cmp            = empty_cmp,
1531         .entry          = node_entry,
1532         .width          = 4,
1533 };
1534
1535 static struct c2c_dimension dim_mean_rmt = {
1536         .header         = HEADER_SPAN("---------- cycles ----------", "rmt hitm", 2),
1537         .name           = "mean_rmt",
1538         .cmp            = empty_cmp,
1539         .entry          = mean_rmt_entry,
1540         .width          = 8,
1541 };
1542
1543 static struct c2c_dimension dim_mean_lcl = {
1544         .header         = HEADER_SPAN_LOW("lcl hitm"),
1545         .name           = "mean_lcl",
1546         .cmp            = empty_cmp,
1547         .entry          = mean_lcl_entry,
1548         .width          = 8,
1549 };
1550
1551 static struct c2c_dimension dim_mean_load = {
1552         .header         = HEADER_SPAN_LOW("load"),
1553         .name           = "mean_load",
1554         .cmp            = empty_cmp,
1555         .entry          = mean_load_entry,
1556         .width          = 8,
1557 };
1558
1559 static struct c2c_dimension dim_cpucnt = {
1560         .header         = HEADER_BOTH("cpu", "cnt"),
1561         .name           = "cpucnt",
1562         .cmp            = empty_cmp,
1563         .entry          = cpucnt_entry,
1564         .width          = 8,
1565 };
1566
1567 static struct c2c_dimension dim_srcline = {
1568         .name           = "cl_srcline",
1569         .se             = &sort_srcline,
1570 };
1571
1572 static struct c2c_dimension dim_dcacheline_idx = {
1573         .header         = HEADER_LOW("Index"),
1574         .name           = "cl_idx",
1575         .cmp            = empty_cmp,
1576         .entry          = cl_idx_entry,
1577         .width          = 5,
1578 };
1579
1580 static struct c2c_dimension dim_dcacheline_num = {
1581         .header         = HEADER_LOW("Num"),
1582         .name           = "cl_num",
1583         .cmp            = empty_cmp,
1584         .entry          = cl_idx_entry,
1585         .width          = 5,
1586 };
1587
1588 static struct c2c_dimension dim_dcacheline_num_empty = {
1589         .header         = HEADER_LOW("Num"),
1590         .name           = "cl_num_empty",
1591         .cmp            = empty_cmp,
1592         .entry          = cl_idx_empty_entry,
1593         .width          = 5,
1594 };
1595
1596 static struct c2c_dimension *dimensions[] = {
1597         &dim_dcacheline,
1598         &dim_dcacheline_node,
1599         &dim_dcacheline_count,
1600         &dim_offset,
1601         &dim_offset_node,
1602         &dim_iaddr,
1603         &dim_tot_hitm,
1604         &dim_lcl_hitm,
1605         &dim_rmt_hitm,
1606         &dim_cl_lcl_hitm,
1607         &dim_cl_rmt_hitm,
1608         &dim_tot_stores,
1609         &dim_stores_l1hit,
1610         &dim_stores_l1miss,
1611         &dim_cl_stores_l1hit,
1612         &dim_cl_stores_l1miss,
1613         &dim_ld_fbhit,
1614         &dim_ld_l1hit,
1615         &dim_ld_l2hit,
1616         &dim_ld_llchit,
1617         &dim_ld_rmthit,
1618         &dim_tot_recs,
1619         &dim_tot_loads,
1620         &dim_percent_hitm,
1621         &dim_percent_rmt_hitm,
1622         &dim_percent_lcl_hitm,
1623         &dim_percent_stores_l1hit,
1624         &dim_percent_stores_l1miss,
1625         &dim_dram_lcl,
1626         &dim_dram_rmt,
1627         &dim_pid,
1628         &dim_tid,
1629         &dim_symbol,
1630         &dim_dso,
1631         &dim_node,
1632         &dim_mean_rmt,
1633         &dim_mean_lcl,
1634         &dim_mean_load,
1635         &dim_cpucnt,
1636         &dim_srcline,
1637         &dim_dcacheline_idx,
1638         &dim_dcacheline_num,
1639         &dim_dcacheline_num_empty,
1640         NULL,
1641 };
1642
1643 static void fmt_free(struct perf_hpp_fmt *fmt)
1644 {
1645         struct c2c_fmt *c2c_fmt;
1646
1647         c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
1648         free(c2c_fmt);
1649 }
1650
1651 static bool fmt_equal(struct perf_hpp_fmt *a, struct perf_hpp_fmt *b)
1652 {
1653         struct c2c_fmt *c2c_a = container_of(a, struct c2c_fmt, fmt);
1654         struct c2c_fmt *c2c_b = container_of(b, struct c2c_fmt, fmt);
1655
1656         return c2c_a->dim == c2c_b->dim;
1657 }
1658
1659 static struct c2c_dimension *get_dimension(const char *name)
1660 {
1661         unsigned int i;
1662
1663         for (i = 0; dimensions[i]; i++) {
1664                 struct c2c_dimension *dim = dimensions[i];
1665
1666                 if (!strcmp(dim->name, name))
1667                         return dim;
1668         }
1669
1670         return NULL;
1671 }
1672
1673 static int c2c_se_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1674                         struct hist_entry *he)
1675 {
1676         struct c2c_fmt *c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
1677         struct c2c_dimension *dim = c2c_fmt->dim;
1678         size_t len = fmt->user_len;
1679
1680         if (!len) {
1681                 len = hists__col_len(he->hists, dim->se->se_width_idx);
1682
1683                 if (dim == &dim_symbol || dim == &dim_srcline)
1684                         len = symbol_width(he->hists, dim->se);
1685         }
1686
1687         return dim->se->se_snprintf(he, hpp->buf, hpp->size, len);
1688 }
1689
1690 static int64_t c2c_se_cmp(struct perf_hpp_fmt *fmt,
1691                           struct hist_entry *a, struct hist_entry *b)
1692 {
1693         struct c2c_fmt *c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
1694         struct c2c_dimension *dim = c2c_fmt->dim;
1695
1696         return dim->se->se_cmp(a, b);
1697 }
1698
1699 static int64_t c2c_se_collapse(struct perf_hpp_fmt *fmt,
1700                                struct hist_entry *a, struct hist_entry *b)
1701 {
1702         struct c2c_fmt *c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
1703         struct c2c_dimension *dim = c2c_fmt->dim;
1704         int64_t (*collapse_fn)(struct hist_entry *, struct hist_entry *);
1705
1706         collapse_fn = dim->se->se_collapse ?: dim->se->se_cmp;
1707         return collapse_fn(a, b);
1708 }
1709
1710 static struct c2c_fmt *get_format(const char *name)
1711 {
1712         struct c2c_dimension *dim = get_dimension(name);
1713         struct c2c_fmt *c2c_fmt;
1714         struct perf_hpp_fmt *fmt;
1715
1716         if (!dim)
1717                 return NULL;
1718
1719         c2c_fmt = zalloc(sizeof(*c2c_fmt));
1720         if (!c2c_fmt)
1721                 return NULL;
1722
1723         c2c_fmt->dim = dim;
1724
1725         fmt = &c2c_fmt->fmt;
1726         INIT_LIST_HEAD(&fmt->list);
1727         INIT_LIST_HEAD(&fmt->sort_list);
1728
1729         fmt->cmp        = dim->se ? c2c_se_cmp   : dim->cmp;
1730         fmt->sort       = dim->se ? c2c_se_cmp   : dim->cmp;
1731         fmt->color      = dim->se ? NULL         : dim->color;
1732         fmt->entry      = dim->se ? c2c_se_entry : dim->entry;
1733         fmt->header     = c2c_header;
1734         fmt->width      = c2c_width;
1735         fmt->collapse   = dim->se ? c2c_se_collapse : dim->cmp;
1736         fmt->equal      = fmt_equal;
1737         fmt->free       = fmt_free;
1738
1739         return c2c_fmt;
1740 }
1741
1742 static int c2c_hists__init_output(struct perf_hpp_list *hpp_list, char *name)
1743 {
1744         struct c2c_fmt *c2c_fmt = get_format(name);
1745
1746         if (!c2c_fmt) {
1747                 reset_dimensions();
1748                 return output_field_add(hpp_list, name);
1749         }
1750
1751         perf_hpp_list__column_register(hpp_list, &c2c_fmt->fmt);
1752         return 0;
1753 }
1754
1755 static int c2c_hists__init_sort(struct perf_hpp_list *hpp_list, char *name)
1756 {
1757         struct c2c_fmt *c2c_fmt = get_format(name);
1758         struct c2c_dimension *dim;
1759
1760         if (!c2c_fmt) {
1761                 reset_dimensions();
1762                 return sort_dimension__add(hpp_list, name, NULL, 0);
1763         }
1764
1765         dim = c2c_fmt->dim;
1766         if (dim == &dim_dso)
1767                 hpp_list->dso = 1;
1768
1769         perf_hpp_list__register_sort_field(hpp_list, &c2c_fmt->fmt);
1770         return 0;
1771 }
1772
1773 #define PARSE_LIST(_list, _fn)                                                  \
1774         do {                                                                    \
1775                 char *tmp, *tok;                                                \
1776                 ret = 0;                                                        \
1777                                                                                 \
1778                 if (!_list)                                                     \
1779                         break;                                                  \
1780                                                                                 \
1781                 for (tok = strtok_r((char *)_list, ", ", &tmp);                 \
1782                                 tok; tok = strtok_r(NULL, ", ", &tmp)) {        \
1783                         ret = _fn(hpp_list, tok);                               \
1784                         if (ret == -EINVAL) {                                   \
1785                                 pr_err("Invalid --fields key: `%s'", tok);      \
1786                                 break;                                          \
1787                         } else if (ret == -ESRCH) {                             \
1788                                 pr_err("Unknown --fields key: `%s'", tok);      \
1789                                 break;                                          \
1790                         }                                                       \
1791                 }                                                               \
1792         } while (0)
1793
1794 static int hpp_list__parse(struct perf_hpp_list *hpp_list,
1795                            const char *output_,
1796                            const char *sort_)
1797 {
1798         char *output = output_ ? strdup(output_) : NULL;
1799         char *sort   = sort_   ? strdup(sort_) : NULL;
1800         int ret;
1801
1802         PARSE_LIST(output, c2c_hists__init_output);
1803         PARSE_LIST(sort,   c2c_hists__init_sort);
1804
1805         /* copy sort keys to output fields */
1806         perf_hpp__setup_output_field(hpp_list);
1807
1808         /*
1809          * We dont need other sorting keys other than those
1810          * we already specified. It also really slows down
1811          * the processing a lot with big number of output
1812          * fields, so switching this off for c2c.
1813          */
1814
1815 #if 0
1816         /* and then copy output fields to sort keys */
1817         perf_hpp__append_sort_keys(&hists->list);
1818 #endif
1819
1820         free(output);
1821         free(sort);
1822         return ret;
1823 }
1824
1825 static int c2c_hists__init(struct c2c_hists *hists,
1826                            const char *sort,
1827                            int nr_header_lines)
1828 {
1829         __hists__init(&hists->hists, &hists->list);
1830
1831         /*
1832          * Initialize only with sort fields, we need to resort
1833          * later anyway, and that's where we add output fields
1834          * as well.
1835          */
1836         perf_hpp_list__init(&hists->list);
1837
1838         /* Overload number of header lines.*/
1839         hists->list.nr_header_lines = nr_header_lines;
1840
1841         return hpp_list__parse(&hists->list, NULL, sort);
1842 }
1843
1844 static int c2c_hists__reinit(struct c2c_hists *c2c_hists,
1845                              const char *output,
1846                              const char *sort)
1847 {
1848         perf_hpp__reset_output_field(&c2c_hists->list);
1849         return hpp_list__parse(&c2c_hists->list, output, sort);
1850 }
1851
1852 #define DISPLAY_LINE_LIMIT  0.001
1853
1854 static bool he__display(struct hist_entry *he, struct c2c_stats *stats)
1855 {
1856         struct c2c_hist_entry *c2c_he;
1857         double ld_dist;
1858
1859         if (c2c.show_all)
1860                 return true;
1861
1862         c2c_he = container_of(he, struct c2c_hist_entry, he);
1863
1864 #define FILTER_HITM(__h)                                                \
1865         if (stats->__h) {                                               \
1866                 ld_dist = ((double)c2c_he->stats.__h / stats->__h);     \
1867                 if (ld_dist < DISPLAY_LINE_LIMIT)                       \
1868                         he->filtered = HIST_FILTER__C2C;                \
1869         } else {                                                        \
1870                 he->filtered = HIST_FILTER__C2C;                        \
1871         }
1872
1873         switch (c2c.display) {
1874         case DISPLAY_LCL:
1875                 FILTER_HITM(lcl_hitm);
1876                 break;
1877         case DISPLAY_RMT:
1878                 FILTER_HITM(rmt_hitm);
1879                 break;
1880         case DISPLAY_TOT:
1881                 FILTER_HITM(tot_hitm);
1882         default:
1883                 break;
1884         }
1885
1886 #undef FILTER_HITM
1887
1888         return he->filtered == 0;
1889 }
1890
1891 static inline int valid_hitm_or_store(struct hist_entry *he)
1892 {
1893         struct c2c_hist_entry *c2c_he;
1894         bool has_hitm;
1895
1896         c2c_he = container_of(he, struct c2c_hist_entry, he);
1897         has_hitm = c2c.display == DISPLAY_TOT ? c2c_he->stats.tot_hitm :
1898                    c2c.display == DISPLAY_LCL ? c2c_he->stats.lcl_hitm :
1899                                                 c2c_he->stats.rmt_hitm;
1900         return has_hitm || c2c_he->stats.store;
1901 }
1902
1903 static void set_node_width(struct c2c_hist_entry *c2c_he, int len)
1904 {
1905         struct c2c_dimension *dim;
1906
1907         dim = &c2c.hists == c2c_he->hists ?
1908               &dim_dcacheline_node : &dim_offset_node;
1909
1910         if (len > dim->width)
1911                 dim->width = len;
1912 }
1913
1914 static int set_nodestr(struct c2c_hist_entry *c2c_he)
1915 {
1916         char buf[30];
1917         int len;
1918
1919         if (c2c_he->nodestr)
1920                 return 0;
1921
1922         if (bitmap_weight(c2c_he->nodeset, c2c.nodes_cnt)) {
1923                 len = bitmap_scnprintf(c2c_he->nodeset, c2c.nodes_cnt,
1924                                       buf, sizeof(buf));
1925         } else {
1926                 len = scnprintf(buf, sizeof(buf), "N/A");
1927         }
1928
1929         set_node_width(c2c_he, len);
1930         c2c_he->nodestr = strdup(buf);
1931         return c2c_he->nodestr ? 0 : -ENOMEM;
1932 }
1933
1934 static void calc_width(struct c2c_hist_entry *c2c_he)
1935 {
1936         struct c2c_hists *c2c_hists;
1937
1938         c2c_hists = container_of(c2c_he->he.hists, struct c2c_hists, hists);
1939         hists__calc_col_len(&c2c_hists->hists, &c2c_he->he);
1940         set_nodestr(c2c_he);
1941 }
1942
1943 static int filter_cb(struct hist_entry *he, void *arg __maybe_unused)
1944 {
1945         struct c2c_hist_entry *c2c_he;
1946
1947         c2c_he = container_of(he, struct c2c_hist_entry, he);
1948
1949         if (c2c.show_src && !he->srcline)
1950                 he->srcline = hist_entry__srcline(he);
1951
1952         calc_width(c2c_he);
1953
1954         if (!valid_hitm_or_store(he))
1955                 he->filtered = HIST_FILTER__C2C;
1956
1957         return 0;
1958 }
1959
1960 static int resort_cl_cb(struct hist_entry *he, void *arg __maybe_unused)
1961 {
1962         struct c2c_hist_entry *c2c_he;
1963         struct c2c_hists *c2c_hists;
1964         bool display = he__display(he, &c2c.hitm_stats);
1965
1966         c2c_he = container_of(he, struct c2c_hist_entry, he);
1967         c2c_hists = c2c_he->hists;
1968
1969         if (display && c2c_hists) {
1970                 static unsigned int idx;
1971
1972                 c2c_he->cacheline_idx = idx++;
1973                 calc_width(c2c_he);
1974
1975                 c2c_hists__reinit(c2c_hists, c2c.cl_output, c2c.cl_resort);
1976
1977                 hists__collapse_resort(&c2c_hists->hists, NULL);
1978                 hists__output_resort_cb(&c2c_hists->hists, NULL, filter_cb);
1979         }
1980
1981         return 0;
1982 }
1983
1984 static void setup_nodes_header(void)
1985 {
1986         dim_node.header = header_node[c2c.node_info];
1987 }
1988
1989 static int setup_nodes(struct perf_session *session)
1990 {
1991         struct numa_node *n;
1992         unsigned long **nodes;
1993         int node, cpu;
1994         int *cpu2node;
1995
1996         if (c2c.node_info > 2)
1997                 c2c.node_info = 2;
1998
1999         c2c.nodes_cnt = session->header.env.nr_numa_nodes;
2000         c2c.cpus_cnt  = session->header.env.nr_cpus_avail;
2001
2002         n = session->header.env.numa_nodes;
2003         if (!n)
2004                 return -EINVAL;
2005
2006         nodes = zalloc(sizeof(unsigned long *) * c2c.nodes_cnt);
2007         if (!nodes)
2008                 return -ENOMEM;
2009
2010         c2c.nodes = nodes;
2011
2012         cpu2node = zalloc(sizeof(int) * c2c.cpus_cnt);
2013         if (!cpu2node)
2014                 return -ENOMEM;
2015
2016         for (cpu = 0; cpu < c2c.cpus_cnt; cpu++)
2017                 cpu2node[cpu] = -1;
2018
2019         c2c.cpu2node = cpu2node;
2020
2021         for (node = 0; node < c2c.nodes_cnt; node++) {
2022                 struct perf_cpu_map *map = n[node].map;
2023                 unsigned long *set;
2024
2025                 set = bitmap_alloc(c2c.cpus_cnt);
2026                 if (!set)
2027                         return -ENOMEM;
2028
2029                 nodes[node] = set;
2030
2031                 /* empty node, skip */
2032                 if (perf_cpu_map__empty(map))
2033                         continue;
2034
2035                 for (cpu = 0; cpu < map->nr; cpu++) {
2036                         set_bit(map->map[cpu], set);
2037
2038                         if (WARN_ONCE(cpu2node[map->map[cpu]] != -1, "node/cpu topology bug"))
2039                                 return -EINVAL;
2040
2041                         cpu2node[map->map[cpu]] = node;
2042                 }
2043         }
2044
2045         setup_nodes_header();
2046         return 0;
2047 }
2048
2049 #define HAS_HITMS(__h) ((__h)->stats.lcl_hitm || (__h)->stats.rmt_hitm)
2050
2051 static int resort_hitm_cb(struct hist_entry *he, void *arg __maybe_unused)
2052 {
2053         struct c2c_hist_entry *c2c_he;
2054         c2c_he = container_of(he, struct c2c_hist_entry, he);
2055
2056         if (HAS_HITMS(c2c_he)) {
2057                 c2c.shared_clines++;
2058                 c2c_add_stats(&c2c.hitm_stats, &c2c_he->stats);
2059         }
2060
2061         return 0;
2062 }
2063
2064 static int hists__iterate_cb(struct hists *hists, hists__resort_cb_t cb)
2065 {
2066         struct rb_node *next = rb_first_cached(&hists->entries);
2067         int ret = 0;
2068
2069         while (next) {
2070                 struct hist_entry *he;
2071
2072                 he = rb_entry(next, struct hist_entry, rb_node);
2073                 ret = cb(he, NULL);
2074                 if (ret)
2075                         break;
2076                 next = rb_next(&he->rb_node);
2077         }
2078
2079         return ret;
2080 }
2081
2082 static void print_c2c__display_stats(FILE *out)
2083 {
2084         int llc_misses;
2085         struct c2c_stats *stats = &c2c.hists.stats;
2086
2087         llc_misses = stats->lcl_dram +
2088                      stats->rmt_dram +
2089                      stats->rmt_hit +
2090                      stats->rmt_hitm;
2091
2092         fprintf(out, "=================================================\n");
2093         fprintf(out, "            Trace Event Information              \n");
2094         fprintf(out, "=================================================\n");
2095         fprintf(out, "  Total records                     : %10d\n", stats->nr_entries);
2096         fprintf(out, "  Locked Load/Store Operations      : %10d\n", stats->locks);
2097         fprintf(out, "  Load Operations                   : %10d\n", stats->load);
2098         fprintf(out, "  Loads - uncacheable               : %10d\n", stats->ld_uncache);
2099         fprintf(out, "  Loads - IO                        : %10d\n", stats->ld_io);
2100         fprintf(out, "  Loads - Miss                      : %10d\n", stats->ld_miss);
2101         fprintf(out, "  Loads - no mapping                : %10d\n", stats->ld_noadrs);
2102         fprintf(out, "  Load Fill Buffer Hit              : %10d\n", stats->ld_fbhit);
2103         fprintf(out, "  Load L1D hit                      : %10d\n", stats->ld_l1hit);
2104         fprintf(out, "  Load L2D hit                      : %10d\n", stats->ld_l2hit);
2105         fprintf(out, "  Load LLC hit                      : %10d\n", stats->ld_llchit + stats->lcl_hitm);
2106         fprintf(out, "  Load Local HITM                   : %10d\n", stats->lcl_hitm);
2107         fprintf(out, "  Load Remote HITM                  : %10d\n", stats->rmt_hitm);
2108         fprintf(out, "  Load Remote HIT                   : %10d\n", stats->rmt_hit);
2109         fprintf(out, "  Load Local DRAM                   : %10d\n", stats->lcl_dram);
2110         fprintf(out, "  Load Remote DRAM                  : %10d\n", stats->rmt_dram);
2111         fprintf(out, "  Load MESI State Exclusive         : %10d\n", stats->ld_excl);
2112         fprintf(out, "  Load MESI State Shared            : %10d\n", stats->ld_shared);
2113         fprintf(out, "  Load LLC Misses                   : %10d\n", llc_misses);
2114         fprintf(out, "  LLC Misses to Local DRAM          : %10.1f%%\n", ((double)stats->lcl_dram/(double)llc_misses) * 100.);
2115         fprintf(out, "  LLC Misses to Remote DRAM         : %10.1f%%\n", ((double)stats->rmt_dram/(double)llc_misses) * 100.);
2116         fprintf(out, "  LLC Misses to Remote cache (HIT)  : %10.1f%%\n", ((double)stats->rmt_hit /(double)llc_misses) * 100.);
2117         fprintf(out, "  LLC Misses to Remote cache (HITM) : %10.1f%%\n", ((double)stats->rmt_hitm/(double)llc_misses) * 100.);
2118         fprintf(out, "  Store Operations                  : %10d\n", stats->store);
2119         fprintf(out, "  Store - uncacheable               : %10d\n", stats->st_uncache);
2120         fprintf(out, "  Store - no mapping                : %10d\n", stats->st_noadrs);
2121         fprintf(out, "  Store L1D Hit                     : %10d\n", stats->st_l1hit);
2122         fprintf(out, "  Store L1D Miss                    : %10d\n", stats->st_l1miss);
2123         fprintf(out, "  No Page Map Rejects               : %10d\n", stats->nomap);
2124         fprintf(out, "  Unable to parse data source       : %10d\n", stats->noparse);
2125 }
2126
2127 static void print_shared_cacheline_info(FILE *out)
2128 {
2129         struct c2c_stats *stats = &c2c.hitm_stats;
2130         int hitm_cnt = stats->lcl_hitm + stats->rmt_hitm;
2131
2132         fprintf(out, "=================================================\n");
2133         fprintf(out, "    Global Shared Cache Line Event Information   \n");
2134         fprintf(out, "=================================================\n");
2135         fprintf(out, "  Total Shared Cache Lines          : %10d\n", c2c.shared_clines);
2136         fprintf(out, "  Load HITs on shared lines         : %10d\n", stats->load);
2137         fprintf(out, "  Fill Buffer Hits on shared lines  : %10d\n", stats->ld_fbhit);
2138         fprintf(out, "  L1D hits on shared lines          : %10d\n", stats->ld_l1hit);
2139         fprintf(out, "  L2D hits on shared lines          : %10d\n", stats->ld_l2hit);
2140         fprintf(out, "  LLC hits on shared lines          : %10d\n", stats->ld_llchit + stats->lcl_hitm);
2141         fprintf(out, "  Locked Access on shared lines     : %10d\n", stats->locks);
2142         fprintf(out, "  Store HITs on shared lines        : %10d\n", stats->store);
2143         fprintf(out, "  Store L1D hits on shared lines    : %10d\n", stats->st_l1hit);
2144         fprintf(out, "  Total Merged records              : %10d\n", hitm_cnt + stats->store);
2145 }
2146
2147 static void print_cacheline(struct c2c_hists *c2c_hists,
2148                             struct hist_entry *he_cl,
2149                             struct perf_hpp_list *hpp_list,
2150                             FILE *out)
2151 {
2152         char bf[1000];
2153         struct perf_hpp hpp = {
2154                 .buf            = bf,
2155                 .size           = 1000,
2156         };
2157         static bool once;
2158
2159         if (!once) {
2160                 hists__fprintf_headers(&c2c_hists->hists, out);
2161                 once = true;
2162         } else {
2163                 fprintf(out, "\n");
2164         }
2165
2166         fprintf(out, "  -------------------------------------------------------------\n");
2167         __hist_entry__snprintf(he_cl, &hpp, hpp_list);
2168         fprintf(out, "%s\n", bf);
2169         fprintf(out, "  -------------------------------------------------------------\n");
2170
2171         hists__fprintf(&c2c_hists->hists, false, 0, 0, 0, out, false);
2172 }
2173
2174 static void print_pareto(FILE *out)
2175 {
2176         struct perf_hpp_list hpp_list;
2177         struct rb_node *nd;
2178         int ret;
2179
2180         perf_hpp_list__init(&hpp_list);
2181         ret = hpp_list__parse(&hpp_list,
2182                                 "cl_num,"
2183                                 "cl_rmt_hitm,"
2184                                 "cl_lcl_hitm,"
2185                                 "cl_stores_l1hit,"
2186                                 "cl_stores_l1miss,"
2187                                 "dcacheline",
2188                                 NULL);
2189
2190         if (WARN_ONCE(ret, "failed to setup sort entries\n"))
2191                 return;
2192
2193         nd = rb_first_cached(&c2c.hists.hists.entries);
2194
2195         for (; nd; nd = rb_next(nd)) {
2196                 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
2197                 struct c2c_hist_entry *c2c_he;
2198
2199                 if (he->filtered)
2200                         continue;
2201
2202                 c2c_he = container_of(he, struct c2c_hist_entry, he);
2203                 print_cacheline(c2c_he->hists, he, &hpp_list, out);
2204         }
2205 }
2206
2207 static void print_c2c_info(FILE *out, struct perf_session *session)
2208 {
2209         struct evlist *evlist = session->evlist;
2210         struct evsel *evsel;
2211         bool first = true;
2212
2213         fprintf(out, "=================================================\n");
2214         fprintf(out, "                 c2c details                     \n");
2215         fprintf(out, "=================================================\n");
2216
2217         evlist__for_each_entry(evlist, evsel) {
2218                 fprintf(out, "%-36s: %s\n", first ? "  Events" : "", evsel__name(evsel));
2219                 first = false;
2220         }
2221         fprintf(out, "  Cachelines sort on                : %s HITMs\n",
2222                 display_str[c2c.display]);
2223         fprintf(out, "  Cacheline data grouping           : %s\n", c2c.cl_sort);
2224 }
2225
2226 static void perf_c2c__hists_fprintf(FILE *out, struct perf_session *session)
2227 {
2228         setup_pager();
2229
2230         print_c2c__display_stats(out);
2231         fprintf(out, "\n");
2232         print_shared_cacheline_info(out);
2233         fprintf(out, "\n");
2234         print_c2c_info(out, session);
2235
2236         if (c2c.stats_only)
2237                 return;
2238
2239         fprintf(out, "\n");
2240         fprintf(out, "=================================================\n");
2241         fprintf(out, "           Shared Data Cache Line Table          \n");
2242         fprintf(out, "=================================================\n");
2243         fprintf(out, "#\n");
2244
2245         hists__fprintf(&c2c.hists.hists, true, 0, 0, 0, stdout, true);
2246
2247         fprintf(out, "\n");
2248         fprintf(out, "=================================================\n");
2249         fprintf(out, "      Shared Cache Line Distribution Pareto      \n");
2250         fprintf(out, "=================================================\n");
2251         fprintf(out, "#\n");
2252
2253         print_pareto(out);
2254 }
2255
2256 #ifdef HAVE_SLANG_SUPPORT
2257 static void c2c_browser__update_nr_entries(struct hist_browser *hb)
2258 {
2259         u64 nr_entries = 0;
2260         struct rb_node *nd = rb_first_cached(&hb->hists->entries);
2261
2262         while (nd) {
2263                 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
2264
2265                 if (!he->filtered)
2266                         nr_entries++;
2267
2268                 nd = rb_next(nd);
2269         }
2270
2271         hb->nr_non_filtered_entries = nr_entries;
2272 }
2273
2274 struct c2c_cacheline_browser {
2275         struct hist_browser      hb;
2276         struct hist_entry       *he;
2277 };
2278
2279 static int
2280 perf_c2c_cacheline_browser__title(struct hist_browser *browser,
2281                                   char *bf, size_t size)
2282 {
2283         struct c2c_cacheline_browser *cl_browser;
2284         struct hist_entry *he;
2285         uint64_t addr = 0;
2286
2287         cl_browser = container_of(browser, struct c2c_cacheline_browser, hb);
2288         he = cl_browser->he;
2289
2290         if (he->mem_info)
2291                 addr = cl_address(he->mem_info->daddr.addr);
2292
2293         scnprintf(bf, size, "Cacheline 0x%lx", addr);
2294         return 0;
2295 }
2296
2297 static struct c2c_cacheline_browser*
2298 c2c_cacheline_browser__new(struct hists *hists, struct hist_entry *he)
2299 {
2300         struct c2c_cacheline_browser *browser;
2301
2302         browser = zalloc(sizeof(*browser));
2303         if (browser) {
2304                 hist_browser__init(&browser->hb, hists);
2305                 browser->hb.c2c_filter  = true;
2306                 browser->hb.title       = perf_c2c_cacheline_browser__title;
2307                 browser->he             = he;
2308         }
2309
2310         return browser;
2311 }
2312
2313 static int perf_c2c__browse_cacheline(struct hist_entry *he)
2314 {
2315         struct c2c_hist_entry *c2c_he;
2316         struct c2c_hists *c2c_hists;
2317         struct c2c_cacheline_browser *cl_browser;
2318         struct hist_browser *browser;
2319         int key = -1;
2320         static const char help[] =
2321         " ENTER         Toggle callchains (if present) \n"
2322         " n             Toggle Node details info \n"
2323         " s             Toggle full length of symbol and source line columns \n"
2324         " q             Return back to cacheline list \n";
2325
2326         if (!he)
2327                 return 0;
2328
2329         /* Display compact version first. */
2330         c2c.symbol_full = false;
2331
2332         c2c_he = container_of(he, struct c2c_hist_entry, he);
2333         c2c_hists = c2c_he->hists;
2334
2335         cl_browser = c2c_cacheline_browser__new(&c2c_hists->hists, he);
2336         if (cl_browser == NULL)
2337                 return -1;
2338
2339         browser = &cl_browser->hb;
2340
2341         /* reset abort key so that it can get Ctrl-C as a key */
2342         SLang_reset_tty();
2343         SLang_init_tty(0, 0, 0);
2344
2345         c2c_browser__update_nr_entries(browser);
2346
2347         while (1) {
2348                 key = hist_browser__run(browser, "? - help", true, 0);
2349
2350                 switch (key) {
2351                 case 's':
2352                         c2c.symbol_full = !c2c.symbol_full;
2353                         break;
2354                 case 'n':
2355                         c2c.node_info = (c2c.node_info + 1) % 3;
2356                         setup_nodes_header();
2357                         break;
2358                 case 'q':
2359                         goto out;
2360                 case '?':
2361                         ui_browser__help_window(&browser->b, help);
2362                         break;
2363                 default:
2364                         break;
2365                 }
2366         }
2367
2368 out:
2369         free(cl_browser);
2370         return 0;
2371 }
2372
2373 static int perf_c2c_browser__title(struct hist_browser *browser,
2374                                    char *bf, size_t size)
2375 {
2376         scnprintf(bf, size,
2377                   "Shared Data Cache Line Table     "
2378                   "(%lu entries, sorted on %s HITMs)",
2379                   browser->nr_non_filtered_entries,
2380                   display_str[c2c.display]);
2381         return 0;
2382 }
2383
2384 static struct hist_browser*
2385 perf_c2c_browser__new(struct hists *hists)
2386 {
2387         struct hist_browser *browser = hist_browser__new(hists);
2388
2389         if (browser) {
2390                 browser->title = perf_c2c_browser__title;
2391                 browser->c2c_filter = true;
2392         }
2393
2394         return browser;
2395 }
2396
2397 static int perf_c2c__hists_browse(struct hists *hists)
2398 {
2399         struct hist_browser *browser;
2400         int key = -1;
2401         static const char help[] =
2402         " d             Display cacheline details \n"
2403         " ENTER         Toggle callchains (if present) \n"
2404         " q             Quit \n";
2405
2406         browser = perf_c2c_browser__new(hists);
2407         if (browser == NULL)
2408                 return -1;
2409
2410         /* reset abort key so that it can get Ctrl-C as a key */
2411         SLang_reset_tty();
2412         SLang_init_tty(0, 0, 0);
2413
2414         c2c_browser__update_nr_entries(browser);
2415
2416         while (1) {
2417                 key = hist_browser__run(browser, "? - help", true, 0);
2418
2419                 switch (key) {
2420                 case 'q':
2421                         goto out;
2422                 case 'd':
2423                         perf_c2c__browse_cacheline(browser->he_selection);
2424                         break;
2425                 case '?':
2426                         ui_browser__help_window(&browser->b, help);
2427                         break;
2428                 default:
2429                         break;
2430                 }
2431         }
2432
2433 out:
2434         hist_browser__delete(browser);
2435         return 0;
2436 }
2437
2438 static void perf_c2c_display(struct perf_session *session)
2439 {
2440         if (use_browser == 0)
2441                 perf_c2c__hists_fprintf(stdout, session);
2442         else
2443                 perf_c2c__hists_browse(&c2c.hists.hists);
2444 }
2445 #else
2446 static void perf_c2c_display(struct perf_session *session)
2447 {
2448         use_browser = 0;
2449         perf_c2c__hists_fprintf(stdout, session);
2450 }
2451 #endif /* HAVE_SLANG_SUPPORT */
2452
2453 static char *fill_line(const char *orig, int len)
2454 {
2455         int i, j, olen = strlen(orig);
2456         char *buf;
2457
2458         buf = zalloc(len + 1);
2459         if (!buf)
2460                 return NULL;
2461
2462         j = len / 2 - olen / 2;
2463
2464         for (i = 0; i < j - 1; i++)
2465                 buf[i] = '-';
2466
2467         buf[i++] = ' ';
2468
2469         strcpy(buf + i, orig);
2470
2471         i += olen;
2472
2473         buf[i++] = ' ';
2474
2475         for (; i < len; i++)
2476                 buf[i] = '-';
2477
2478         return buf;
2479 }
2480
2481 static int ui_quirks(void)
2482 {
2483         const char *nodestr = "Data address";
2484         char *buf;
2485
2486         if (!c2c.use_stdio) {
2487                 dim_offset.width  = 5;
2488                 dim_offset.header = header_offset_tui;
2489                 nodestr = "CL";
2490         }
2491
2492         dim_percent_hitm.header = percent_hitm_header[c2c.display];
2493
2494         /* Fix the zero line for dcacheline column. */
2495         buf = fill_line("Cacheline", dim_dcacheline.width +
2496                                      dim_dcacheline_node.width +
2497                                      dim_dcacheline_count.width + 4);
2498         if (!buf)
2499                 return -ENOMEM;
2500
2501         dim_dcacheline.header.line[0].text = buf;
2502
2503         /* Fix the zero line for offset column. */
2504         buf = fill_line(nodestr, dim_offset.width +
2505                                  dim_offset_node.width +
2506                                  dim_dcacheline_count.width + 4);
2507         if (!buf)
2508                 return -ENOMEM;
2509
2510         dim_offset.header.line[0].text = buf;
2511
2512         return 0;
2513 }
2514
2515 #define CALLCHAIN_DEFAULT_OPT  "graph,0.5,caller,function,percent"
2516
2517 const char callchain_help[] = "Display call graph (stack chain/backtrace):\n\n"
2518                                 CALLCHAIN_REPORT_HELP
2519                                 "\n\t\t\t\tDefault: " CALLCHAIN_DEFAULT_OPT;
2520
2521 static int
2522 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
2523 {
2524         struct callchain_param *callchain = opt->value;
2525
2526         callchain->enabled = !unset;
2527         /*
2528          * --no-call-graph
2529          */
2530         if (unset) {
2531                 symbol_conf.use_callchain = false;
2532                 callchain->mode = CHAIN_NONE;
2533                 return 0;
2534         }
2535
2536         return parse_callchain_report_opt(arg);
2537 }
2538
2539 static int setup_callchain(struct evlist *evlist)
2540 {
2541         u64 sample_type = evlist__combined_sample_type(evlist);
2542         enum perf_call_graph_mode mode = CALLCHAIN_NONE;
2543
2544         if ((sample_type & PERF_SAMPLE_REGS_USER) &&
2545             (sample_type & PERF_SAMPLE_STACK_USER)) {
2546                 mode = CALLCHAIN_DWARF;
2547                 dwarf_callchain_users = true;
2548         } else if (sample_type & PERF_SAMPLE_BRANCH_STACK)
2549                 mode = CALLCHAIN_LBR;
2550         else if (sample_type & PERF_SAMPLE_CALLCHAIN)
2551                 mode = CALLCHAIN_FP;
2552
2553         if (!callchain_param.enabled &&
2554             callchain_param.mode != CHAIN_NONE &&
2555             mode != CALLCHAIN_NONE) {
2556                 symbol_conf.use_callchain = true;
2557                 if (callchain_register_param(&callchain_param) < 0) {
2558                         ui__error("Can't register callchain params.\n");
2559                         return -EINVAL;
2560                 }
2561         }
2562
2563         if (c2c.stitch_lbr && (mode != CALLCHAIN_LBR)) {
2564                 ui__warning("Can't find LBR callchain. Switch off --stitch-lbr.\n"
2565                             "Please apply --call-graph lbr when recording.\n");
2566                 c2c.stitch_lbr = false;
2567         }
2568
2569         callchain_param.record_mode = mode;
2570         callchain_param.min_percent = 0;
2571         return 0;
2572 }
2573
2574 static int setup_display(const char *str)
2575 {
2576         const char *display = str ?: "tot";
2577
2578         if (!strcmp(display, "tot"))
2579                 c2c.display = DISPLAY_TOT;
2580         else if (!strcmp(display, "rmt"))
2581                 c2c.display = DISPLAY_RMT;
2582         else if (!strcmp(display, "lcl"))
2583                 c2c.display = DISPLAY_LCL;
2584         else {
2585                 pr_err("failed: unknown display type: %s\n", str);
2586                 return -1;
2587         }
2588
2589         return 0;
2590 }
2591
2592 #define for_each_token(__tok, __buf, __sep, __tmp)              \
2593         for (__tok = strtok_r(__buf, __sep, &__tmp); __tok;     \
2594              __tok = strtok_r(NULL,  __sep, &__tmp))
2595
2596 static int build_cl_output(char *cl_sort, bool no_source)
2597 {
2598         char *tok, *tmp, *buf = strdup(cl_sort);
2599         bool add_pid   = false;
2600         bool add_tid   = false;
2601         bool add_iaddr = false;
2602         bool add_sym   = false;
2603         bool add_dso   = false;
2604         bool add_src   = false;
2605         int ret = 0;
2606
2607         if (!buf)
2608                 return -ENOMEM;
2609
2610         for_each_token(tok, buf, ",", tmp) {
2611                 if (!strcmp(tok, "tid")) {
2612                         add_tid = true;
2613                 } else if (!strcmp(tok, "pid")) {
2614                         add_pid = true;
2615                 } else if (!strcmp(tok, "iaddr")) {
2616                         add_iaddr = true;
2617                         add_sym   = true;
2618                         add_dso   = true;
2619                         add_src   = no_source ? false : true;
2620                 } else if (!strcmp(tok, "dso")) {
2621                         add_dso = true;
2622                 } else if (strcmp(tok, "offset")) {
2623                         pr_err("unrecognized sort token: %s\n", tok);
2624                         ret = -EINVAL;
2625                         goto err;
2626                 }
2627         }
2628
2629         if (asprintf(&c2c.cl_output,
2630                 "%s%s%s%s%s%s%s%s%s%s",
2631                 c2c.use_stdio ? "cl_num_empty," : "",
2632                 "percent_rmt_hitm,"
2633                 "percent_lcl_hitm,"
2634                 "percent_stores_l1hit,"
2635                 "percent_stores_l1miss,"
2636                 "offset,offset_node,dcacheline_count,",
2637                 add_pid   ? "pid," : "",
2638                 add_tid   ? "tid," : "",
2639                 add_iaddr ? "iaddr," : "",
2640                 "mean_rmt,"
2641                 "mean_lcl,"
2642                 "mean_load,"
2643                 "tot_recs,"
2644                 "cpucnt,",
2645                 add_sym ? "symbol," : "",
2646                 add_dso ? "dso," : "",
2647                 add_src ? "cl_srcline," : "",
2648                 "node") < 0) {
2649                 ret = -ENOMEM;
2650                 goto err;
2651         }
2652
2653         c2c.show_src = add_src;
2654 err:
2655         free(buf);
2656         return ret;
2657 }
2658
2659 static int setup_coalesce(const char *coalesce, bool no_source)
2660 {
2661         const char *c = coalesce ?: coalesce_default;
2662
2663         if (asprintf(&c2c.cl_sort, "offset,%s", c) < 0)
2664                 return -ENOMEM;
2665
2666         if (build_cl_output(c2c.cl_sort, no_source))
2667                 return -1;
2668
2669         if (asprintf(&c2c.cl_resort, "offset,%s",
2670                      c2c.display == DISPLAY_TOT ?
2671                      "tot_hitm" :
2672                      c2c.display == DISPLAY_RMT ?
2673                      "rmt_hitm,lcl_hitm" :
2674                      "lcl_hitm,rmt_hitm") < 0)
2675                 return -ENOMEM;
2676
2677         pr_debug("coalesce sort   fields: %s\n", c2c.cl_sort);
2678         pr_debug("coalesce resort fields: %s\n", c2c.cl_resort);
2679         pr_debug("coalesce output fields: %s\n", c2c.cl_output);
2680         return 0;
2681 }
2682
2683 static int perf_c2c__report(int argc, const char **argv)
2684 {
2685         struct itrace_synth_opts itrace_synth_opts = {
2686                 .set = true,
2687                 .mem = true,    /* Only enable memory event */
2688                 .default_no_sample = true,
2689         };
2690
2691         struct perf_session *session;
2692         struct ui_progress prog;
2693         struct perf_data data = {
2694                 .mode = PERF_DATA_MODE_READ,
2695         };
2696         char callchain_default_opt[] = CALLCHAIN_DEFAULT_OPT;
2697         const char *display = NULL;
2698         const char *coalesce = NULL;
2699         bool no_source = false;
2700         const struct option options[] = {
2701         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
2702                    "file", "vmlinux pathname"),
2703         OPT_STRING('i', "input", &input_name, "file",
2704                    "the input file to process"),
2705         OPT_INCR('N', "node-info", &c2c.node_info,
2706                  "show extra node info in report (repeat for more info)"),
2707 #ifdef HAVE_SLANG_SUPPORT
2708         OPT_BOOLEAN(0, "stdio", &c2c.use_stdio, "Use the stdio interface"),
2709 #endif
2710         OPT_BOOLEAN(0, "stats", &c2c.stats_only,
2711                     "Display only statistic tables (implies --stdio)"),
2712         OPT_BOOLEAN(0, "full-symbols", &c2c.symbol_full,
2713                     "Display full length of symbols"),
2714         OPT_BOOLEAN(0, "no-source", &no_source,
2715                     "Do not display Source Line column"),
2716         OPT_BOOLEAN(0, "show-all", &c2c.show_all,
2717                     "Show all captured HITM lines."),
2718         OPT_CALLBACK_DEFAULT('g', "call-graph", &callchain_param,
2719                              "print_type,threshold[,print_limit],order,sort_key[,branch],value",
2720                              callchain_help, &parse_callchain_opt,
2721                              callchain_default_opt),
2722         OPT_STRING('d', "display", &display, "Switch HITM output type", "lcl,rmt"),
2723         OPT_STRING('c', "coalesce", &coalesce, "coalesce fields",
2724                    "coalesce fields: pid,tid,iaddr,dso"),
2725         OPT_BOOLEAN('f', "force", &symbol_conf.force, "don't complain, do it"),
2726         OPT_BOOLEAN(0, "stitch-lbr", &c2c.stitch_lbr,
2727                     "Enable LBR callgraph stitching approach"),
2728         OPT_PARENT(c2c_options),
2729         OPT_END()
2730         };
2731         int err = 0;
2732
2733         argc = parse_options(argc, argv, options, report_c2c_usage,
2734                              PARSE_OPT_STOP_AT_NON_OPTION);
2735         if (argc)
2736                 usage_with_options(report_c2c_usage, options);
2737
2738         if (c2c.stats_only)
2739                 c2c.use_stdio = true;
2740
2741         if (!input_name || !strlen(input_name))
2742                 input_name = "perf.data";
2743
2744         data.path  = input_name;
2745         data.force = symbol_conf.force;
2746
2747         err = setup_display(display);
2748         if (err)
2749                 goto out;
2750
2751         err = setup_coalesce(coalesce, no_source);
2752         if (err) {
2753                 pr_debug("Failed to initialize hists\n");
2754                 goto out;
2755         }
2756
2757         err = c2c_hists__init(&c2c.hists, "dcacheline", 2);
2758         if (err) {
2759                 pr_debug("Failed to initialize hists\n");
2760                 goto out;
2761         }
2762
2763         session = perf_session__new(&data, 0, &c2c.tool);
2764         if (IS_ERR(session)) {
2765                 err = PTR_ERR(session);
2766                 pr_debug("Error creating perf session\n");
2767                 goto out;
2768         }
2769
2770         session->itrace_synth_opts = &itrace_synth_opts;
2771
2772         err = setup_nodes(session);
2773         if (err) {
2774                 pr_err("Failed setup nodes\n");
2775                 goto out;
2776         }
2777
2778         err = mem2node__init(&c2c.mem2node, &session->header.env);
2779         if (err)
2780                 goto out_session;
2781
2782         err = setup_callchain(session->evlist);
2783         if (err)
2784                 goto out_mem2node;
2785
2786         if (symbol__init(&session->header.env) < 0)
2787                 goto out_mem2node;
2788
2789         /* No pipe support at the moment. */
2790         if (perf_data__is_pipe(session->data)) {
2791                 pr_debug("No pipe support at the moment.\n");
2792                 goto out_mem2node;
2793         }
2794
2795         if (c2c.use_stdio)
2796                 use_browser = 0;
2797         else
2798                 use_browser = 1;
2799
2800         setup_browser(false);
2801
2802         err = perf_session__process_events(session);
2803         if (err) {
2804                 pr_err("failed to process sample\n");
2805                 goto out_mem2node;
2806         }
2807
2808         c2c_hists__reinit(&c2c.hists,
2809                         "cl_idx,"
2810                         "dcacheline,"
2811                         "dcacheline_node,"
2812                         "dcacheline_count,"
2813                         "percent_hitm,"
2814                         "tot_hitm,lcl_hitm,rmt_hitm,"
2815                         "tot_recs,"
2816                         "tot_loads,"
2817                         "tot_stores,"
2818                         "stores_l1hit,stores_l1miss,"
2819                         "ld_fbhit,ld_l1hit,ld_l2hit,"
2820                         "ld_lclhit,lcl_hitm,"
2821                         "ld_rmthit,rmt_hitm,"
2822                         "dram_lcl,dram_rmt",
2823                         c2c.display == DISPLAY_TOT ? "tot_hitm" :
2824                         c2c.display == DISPLAY_LCL ? "lcl_hitm" : "rmt_hitm"
2825                         );
2826
2827         ui_progress__init(&prog, c2c.hists.hists.nr_entries, "Sorting...");
2828
2829         hists__collapse_resort(&c2c.hists.hists, NULL);
2830         hists__output_resort_cb(&c2c.hists.hists, &prog, resort_hitm_cb);
2831         hists__iterate_cb(&c2c.hists.hists, resort_cl_cb);
2832
2833         ui_progress__finish();
2834
2835         if (ui_quirks()) {
2836                 pr_err("failed to setup UI\n");
2837                 goto out_mem2node;
2838         }
2839
2840         perf_c2c_display(session);
2841
2842 out_mem2node:
2843         mem2node__exit(&c2c.mem2node);
2844 out_session:
2845         perf_session__delete(session);
2846 out:
2847         return err;
2848 }
2849
2850 static int parse_record_events(const struct option *opt,
2851                                const char *str, int unset __maybe_unused)
2852 {
2853         bool *event_set = (bool *) opt->value;
2854
2855         if (!strcmp(str, "list")) {
2856                 perf_mem_events__list();
2857                 exit(0);
2858         }
2859         if (perf_mem_events__parse(str))
2860                 exit(-1);
2861
2862         *event_set = true;
2863         return 0;
2864 }
2865
2866
2867 static const char * const __usage_record[] = {
2868         "perf c2c record [<options>] [<command>]",
2869         "perf c2c record [<options>] -- <command> [<options>]",
2870         NULL
2871 };
2872
2873 static const char * const *record_mem_usage = __usage_record;
2874
2875 static int perf_c2c__record(int argc, const char **argv)
2876 {
2877         int rec_argc, i = 0, j;
2878         const char **rec_argv;
2879         int ret;
2880         bool all_user = false, all_kernel = false;
2881         bool event_set = false;
2882         struct perf_mem_event *e;
2883         struct option options[] = {
2884         OPT_CALLBACK('e', "event", &event_set, "event",
2885                      "event selector. Use 'perf c2c record -e list' to list available events",
2886                      parse_record_events),
2887         OPT_BOOLEAN('u', "all-user", &all_user, "collect only user level data"),
2888         OPT_BOOLEAN('k', "all-kernel", &all_kernel, "collect only kernel level data"),
2889         OPT_UINTEGER('l', "ldlat", &perf_mem_events__loads_ldlat, "setup mem-loads latency"),
2890         OPT_PARENT(c2c_options),
2891         OPT_END()
2892         };
2893
2894         if (perf_mem_events__init()) {
2895                 pr_err("failed: memory events not supported\n");
2896                 return -1;
2897         }
2898
2899         argc = parse_options(argc, argv, options, record_mem_usage,
2900                              PARSE_OPT_KEEP_UNKNOWN);
2901
2902         rec_argc = argc + 11; /* max number of arguments */
2903         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2904         if (!rec_argv)
2905                 return -1;
2906
2907         rec_argv[i++] = "record";
2908
2909         if (!event_set) {
2910                 e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD_STORE);
2911                 /*
2912                  * The load and store operations are required, use the event
2913                  * PERF_MEM_EVENTS__LOAD_STORE if it is supported.
2914                  */
2915                 if (e->tag) {
2916                         e->record = true;
2917                 } else {
2918                         e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD);
2919                         e->record = true;
2920
2921                         e = perf_mem_events__ptr(PERF_MEM_EVENTS__STORE);
2922                         e->record = true;
2923                 }
2924         }
2925
2926         e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD);
2927         if (e->record)
2928                 rec_argv[i++] = "-W";
2929
2930         rec_argv[i++] = "-d";
2931         rec_argv[i++] = "--phys-data";
2932         rec_argv[i++] = "--sample-cpu";
2933
2934         for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
2935                 e = perf_mem_events__ptr(j);
2936                 if (!e->record)
2937                         continue;
2938
2939                 if (!e->supported) {
2940                         pr_err("failed: event '%s' not supported\n",
2941                                perf_mem_events__name(j));
2942                         free(rec_argv);
2943                         return -1;
2944                 }
2945
2946                 rec_argv[i++] = "-e";
2947                 rec_argv[i++] = perf_mem_events__name(j);
2948         }
2949
2950         if (all_user)
2951                 rec_argv[i++] = "--all-user";
2952
2953         if (all_kernel)
2954                 rec_argv[i++] = "--all-kernel";
2955
2956         for (j = 0; j < argc; j++, i++)
2957                 rec_argv[i] = argv[j];
2958
2959         if (verbose > 0) {
2960                 pr_debug("calling: ");
2961
2962                 j = 0;
2963
2964                 while (rec_argv[j]) {
2965                         pr_debug("%s ", rec_argv[j]);
2966                         j++;
2967                 }
2968                 pr_debug("\n");
2969         }
2970
2971         ret = cmd_record(i, rec_argv);
2972         free(rec_argv);
2973         return ret;
2974 }
2975
2976 int cmd_c2c(int argc, const char **argv)
2977 {
2978         argc = parse_options(argc, argv, c2c_options, c2c_usage,
2979                              PARSE_OPT_STOP_AT_NON_OPTION);
2980
2981         if (!argc)
2982                 usage_with_options(c2c_usage, c2c_options);
2983
2984         if (!strncmp(argv[0], "rec", 3)) {
2985                 return perf_c2c__record(argc, argv);
2986         } else if (!strncmp(argv[0], "rep", 3)) {
2987                 return perf_c2c__report(argc, argv);
2988         } else {
2989                 usage_with_options(c2c_usage, c2c_options);
2990         }
2991
2992         return 0;
2993 }